Pith.

Computation and Language · arXiv:2611.26358 v1 · submitted 19 Nov 2026 · 199 references

Beam Model Technical Report

Reflection Team: Adam Abdulhamid, Osama Abuelsorour, Yasmin Alameddine, Carlo Alfano, Duarte Alves, Brandon Amos, Selvaraj Anandaraj, Arseni Anisimovich

show 221 more authors

George Antoniadis, Ioannis Antonoglou, Sheila Babayan, Neeraj Bahl, Märt Bakler, Vidhisha Balachandran, Cameron Banks, Thibaut Barroyer, Navroop Bath, Irwan Bello, Jeremy Berman, Fernanda Bernaldo, Mrityunjay Bhadauria, Sanyam Bhutani, Tolga Bolukbasi, Anudhyan Boral, Joseph Brennan, Ethan Brooks, Sam Buchanan, Anton Bulyenov, John Cai, Catherine Cang, Eason Chan, Elijah Chanakira, Nathan Chandrasekaran, Hao Chen, Hong-You Chen, Jeffrey Chen, Saurav Chhatrapati, Ting-Rui Chiang, Brian Tiger Chow, Aakanksha Chowdhery, Timothy Chung, Rob Claire, John D. Co-Reyes, Anthony Conte, Eduardo Coronado-Montoya, Julien Crozon, Drew Csillag, Joe Cummings, Jeff Da, Samira Daruki, Tim Dawson, Charlie Deck, Arash Delijani, Ousmane Dia, Randy Dodgen, Ryan Douglas, Omar El-Kishky, Ben Ellis, Gamaleldin Elsayed, Simon Fan, Patrick Fernandes, Gabriel Ferns, Nick Fram, Ankush Garg, Florent Gbelidji, Iaroslav Geraskin, Behrooz Ghorbani, Ivan Glushkov, Neil Gollapudi, Scott Goodfriend, Yoni Gozlan, Christopher Grimm, Suriya Gunasekar, Shruti Gupta, Joel Gustafson, Emre Guven, Jon Hackmann, Jeremy Hadidjojo, Jess Hamrick, Ethan Hellman, Joscha Hoche, Spurthi Amba Hombaiah, Uzair Honda, Zijian Hu, Dzmitry Huba, Nancy Hung, Vladimir Ivanov, Shomik Jain, Mojan Javaheripi, Kevin Martin Jose, Aymen Kallala, Amog Kamsetty, Min Hyung Kang, Steven Kapturowski, Ramtin Keramati, Samuel Kessler, Sabina Kim, Seonghyeon Kim, Ivan Komarov, Agni Kumar, Thanard Kurutach, Jessica Landon, James Larochelle, Michael Laskin, Jonathan Lee, Jonathan Leung, Chenyu Li, Haoran Li, Rongzhao Li, Michelle Light, Boyan Lin, Hai Lin, Xiayue Charles Lin, Michael Linden, Ollie Liu, Wei Liu, Zichang Liu, Brian Lui, Amr Mahdi, Raunak Manjani, Kobe Martin, Lucas Maystre, Francis McCann, Brendan McLaughlin, Isabel Mehta, Alishah Momin, Gerard Moreno-Torres Bertran, Adriana Murga, Swarnashree Mysore Sathyendra, Ashley Nagaki, Vishal Nandavanam, Swaroop Nath, Jake Nations, Angus Ng, Thien Nguyen, Tyler Nichols, Anaelia Ovalle, James Pelly, Michelle Pieszko, Sashank Pisupati, Oleksandr Polozov, Piyush Prasad, Drew Proud, Salman Quazi, Asfandyar Qureshi, Marko Radmilac, Mohit Raghavendra, Vikash Rai, Ralph Recto, Anurag Reddy, Aksel Reedi, Abby Reider, David Renardy, Matthias Reso, Pierre Roux, David Rubinstein, Paul Ruh, Ildar Safarov, Himanshu Sahni, Adrian Samatan, Simon Sandrew, Adriel Saporta, Rohan Savargaonkar, Vlad Savinov, Nick Schank, Jonathon Schwartz, Tim Seah, Ahmad Sharif, Miriam Shlimovich, Hamid Shojanazeri, Yusuf Sindi, Teak Sittipunt, Kurt Smith, Nicolas Sonnerat, Joe Spisak, Raunak Srivastava, Richard Steigerwald, Adam Stein, Stu Stewart, Arnaud Stiegler, Jingtong Su, Alicia Sun, Zoey Sun, Jakub Sygnowski, Douglas Teixeira, Chetan Tekur, Zachary Tellman, Helen Thatcher, Romal Thoppilan, Anil Tolwani, Konstantin Topaloglou-Mundy, Dmitry Ulyumdzhiev, Meet Vadakkanchery, Christopher Vanderschuere, Balaji Varadarajan, Rohan Varma, Balaji Veeramani, Andreas Verdelis, Risto Vuorio, Tom Walshe, Allen Wang, Ewen Wang, Taiyi Wang, Zach Wentz, Peter Wiese, Travis Wilson, Jason Winters, Sam Wiseman, Arvin Wu, Jin Wu, Kai Wu, Kenneth Xu, Mira Yadav, Xianjun Yang, Yao-Yuan Yang, Xiaolan You, Renault Young, Christina Yu, Biao Zhang, Gavin Zhang, Mingrui Zhang, Shenao Zhang, Jiesi Zhao, Mengjun Zhao, Xingchen Zhao, Mike Zhou, Tong Zhou, Zhenpeng Zhou, Tina Zhu

The paper says

“Beam is a sparse Mixture-of-Experts (MoE) model with 501.3B total parameters and 23.4B active parameters, specifically built and optimized for coding, reasoning, and agentic workloads.”

more of the abstract

Beam delivers frontier level reasoning, while requiring 3-4x less inference compute than many models of its class. Pretrained on 23.8T diverse, curated, high-quality tokens, Beam features a 1M-token context window achieved through a novel two-stage midtraining phase, built on verifiable, correctness-anchored data. Post-training leveraged a large-scale, fully asynchronous reinforcement learning (RL) campaign across 10.5K GB300 GPUs, generating over 100M rollouts over 4 weeks of training. We believe this is one of the largest scale RL runs conducted by any open lab to date. Across our evaluation suite, capabilities continued to improve as we increased RL compute, with no sign of a plateau. Supported by deep hardware-model co-design, FP8/NVFP4 quantization, and a 1.93B-parameter speculative decoding drafter (DFlash 2), Beam maximizes serving throughput with efficient, flexible deployment. Averaged across conversation turns and harm categories, Beam refuses or pivots on 88.9% of harmful requests while complying with 81.5% of harmless requests, achieving the highest refusal/pivot rate on harmful requests among open-weight models on the Pareto frontier.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 20 Nov 2026 · read 599,600 of 609,632 characters · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions.

18 minor · 1 open question

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

The paper, first character to last

Not given to the reviewer: 10,032 of the paper's 609,632 characters. Not checked: figures, code, what the 199 references say.

18 minor objections

The paper says, at §1 (Introduction)

Pith says: minor objection 1 of 18

The phrase 'while delivering frontier-level reasoning, while requiring 3-4x less inference compute' is redundant and grammatically awkward. Consider rephrasing for clarity.

The paper says, at §2.1.1 (Model Architecture, Table 1)

Pith says: minor objection 2 of 18

The table lists 'Total Parameters' as 501.3B and 'Activated Parameters' as 23.4B. It would be helpful to also include the number of parameters for the embedding layer and the LM head, as these are often excluded from 'non-embedding' parameter counts used in scaling laws.

show 16 more

The paper says, at §2.1.2 (Training Recipe, Table 2)

Pith says: minor objection 3 of 18

The peak learning rate is given with many decimal places (7.63701912×10^{-4}). It is unclear whether this level of precision is meaningful or an artifact of the tuning process. The authors should comment on the sensitivity of the training to small changes in the learning rate.

The paper says, at §2.2.1 (Web Data Curation, Figure 2)

Pith says: minor objection 4 of 18

The figure is described but not visible in the supplied text. The caption mentions 'domain-specific filtering and quality partitioning' but does not specify the criteria used for partitioning. The authors should ensure the figure clearly illustrates the decision boundaries and thresholds used.

The paper says, at §2.2.3 (STEM and PDF Data Curation)

Pith says: minor objection 5 of 18

The paper does not specify which OCR model was used. This is important for reproducibility. The authors should either name the model or provide sufficient detail about its architecture and training data.

The paper says, at §3.1 (Optimization)

Pith says: minor objection 6 of 18

The paper states that 5% of the maximum pretraining learning rate was found to be optimal, but does not provide details on the sweep range or the number of configurations tested. This makes it difficult to assess the robustness of this finding.

The paper says, at §4.2.2 (Asynchronous rollouts)

Pith says: minor objection 7 of 18

The paper mentions that inference workers are updated after every optimizer step, but does not specify the mechanism for this update (e.g., whether it involves broadcasting weights or using a parameter server). This is important for understanding the system's scalability and potential bottlenecks.

The paper says, at §4.2.2 (Reward integrity)

Pith says: minor objection 8 of 18

The paper mentions a 'reward-hacking rubric' but does not provide details on its construction or validation. Given the importance of reward integrity for RL, the authors should describe how the rubric was developed and tested.

The paper says, at §4.3.2 (SFT Data Mixture, Table 5)

Pith says: minor objection 9 of 18

The table shows that 'Safety: direct requests + jailbreak/over-refusal' constitutes only 1.62% of the SFT mixture by trace and 0.18% by tokens. This seems very small for a component that is critical for safety. The authors should explain why this proportion is sufficient and how it was determined.

The paper says, at §4.3.3 (RL Training, Knowledge Calibration)

Pith says: minor objection 10 of 18

Similar to the previous comment, the paper should justify why 3% is the right proportion for knowledge calibration in the RL mixture, and discuss how this was determined experimentally.

The paper says, at §5 (Evaluations, Table 8)

Pith says: minor objection 11 of 18

The table includes many benchmarks but does not provide confidence intervals or error bars. Given the stochastic nature of LLM evaluations, especially with temperature sampling, the authors should report uncertainty estimates (e.g., bootstrap confidence intervals) to allow meaningful comparisons between models.

The paper says, at §5 (Evaluations, Table 9)

Pith says: minor objection 12 of 18

The paper reports HLE (no tools) scores but does not specify the evaluation protocol (e.g., number of samples, temperature). This is important for reproducibility. The authors should provide these details in the main text or in an appendix.

The paper says, at §6 (Infrastructure, Figure 14)

Pith says: minor objection 13 of 18

The figure is described but not visible. The caption mentions 'logical interactions rather than transport protocols.' The authors should ensure the figure is clear and includes all components mentioned in the text (control service, rollout workers, training service, etc.).

The paper says, at §6.4.2 (Goodput, Figure 16)

Pith says: minor objection 14 of 18

The figure is not visible. The caption mentions 'operational events' taking 1.7 days. The authors should clarify what constitutes an 'operational event' and whether this time is included in the goodput calculation.

The paper says, at §7 (Safety, Figure 18)

Pith says: minor objection 15 of 18

The figure is not visible. The caption mentions 'observed safety-helpfulness Pareto frontier.' The authors should define what models are included in the 'observed' set and how the frontier is constructed.

The paper says, at §7.3.2 (Conversational ISE, Table 14)

Pith says: minor objection 16 of 18

The table shows results for many models but does not include confidence intervals. Given the relatively small number of prompts per category (100 conversations per category), the authors should report uncertainty estimates to assess the reliability of the comparisons.

The paper says, at §7.3.3 (Agentic ISE, Table 17)

Pith says: minor objection 17 of 18

The paper reports harmful compliance rates but does not specify the number of scenarios or rollouts per scenario. This information is necessary to assess the statistical significance of the reported differences between models.

The paper says, at §A (Hyperparameter Tuning)

Pith says: minor objection 18 of 18

The appendix describes a method for tuning the effective learning rate using a quadratic fit. However, it does not report the goodness of fit or the number of data points used. The authors should provide these details to allow readers to assess the reliability of the tuning procedure.

Open questions (1)

Pith raised these and could not confirm them against the text of the paper it read. They are questions for the authors, not findings.

  1. §4.2.2 (Generalization) Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.

What would settle it

Train a model with the same architecture, data, and RL recipe but with a different random seed or on a different hardware configuration. If the RL training plateaus or shows diminishing returns earlier than reported, or if the model fails to achieve comparable benchmark scores, the central claim of predictable scaling and continued improvement would be undermined.

from Pith's summary of the paper

How the review went

4 passes by deepseek-v4-flash · 20 Nov 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic concern withdrawn: Pith's reviewer misread the paper 35 s · 153,252 tokens read, 5,647 written
  3. Referee 18 minor 34 s · 153,663 tokens read, 2,913 written
  4. Check against the paper kept 0 of 6 objections 1 min 2 s · 309,170 tokens read, 6,953 written

Not run on this paper: the editor, the simulated authors' defence.

Pith's passes raised 6 objections. A check against the paper's text kept 0, withdrew 5 (5 Pith's reviewer misread the paper) and left 1 as a question.

The review in full

The stress test's concern

Withdrawn: Pith's reviewer misread the paper.

The paper says

“Beam delivers frontier level reasoning, while requiring 3-4x less inference compute than many models of its class.”

The referee's report

Not published. Pith's referee wrote this assessment before its objections were checked against the paper, and 5 of its 5 major objections did not survive that check.

Pith's reading of the paper (a machine's summary, not the authors')

The paper introduces Beam, an open-weight sparse Mixture-of-Experts (MoE) language model with 501.3B total parameters and 23.4B active parameters, developed by Reflection. The authors claim that Beam delivers frontier-level reasoning performance on coding, mathematics, and agentic tasks while requiring 3-4x less inference compute than comparable models. This efficiency is achieved through a combination of novel architectural components (depth-scaled residual connections, adaptive expert bias updates), rigorous data curation pipelines (domain-specific filtering, quality-aware upsampling, large-scale OCR), a two-stage midtraining pipeline that extends context length to 1M tokens, and a massive asynchronous reinforcement learning campaign that generated over 100M rollouts on 10.5K GPUs. The paper also reports that Beam achieves the highest refusal/pivot rate on harmful requests among open-weight models on the Pareto frontier (88.9% harmful refusal, 81.5% harmless compliance). If true, this would represent a significant advance in open-weight model capabilities, particularly for agentic and reasoning-heavy workloads, while maintaining strong safety alignment.

What carries the argument. The central object is the depth-scaled residual connection (DSRes), which rescales the residual stream by 1 over the square root of the number of accumulated sublayers: $\mathbf{x}_n = \frac{1}{\sqrt{n}} \sum_{t=1}^n \mathbf{h}_t$. This ensures that the RMS of the residual stream remains approximately constant across depth, preventing unbounded activation growth. The adaptive expert bias update schedule uses a cosine decay from a maximum learning rate, improving routing stability at scale. The two-stage midtraining pipeline first develops core reasoning capabilities at 32K context, then extends to 512K context using a mixture of organic and synthetic long-context data. The asynchronous RL system uses GRPO-style policy gradients with routing replay, nucleus sampling consistency, and task-dependent rollout deadlines to handle heavy-tailed rollout durations.

Core claim. The central claim is that Beam, a 501.3B-parameter sparse MoE model (23.4B active parameters), achieves frontier-level reasoning and agentic capabilities while being 3-4x more inference-efficient than many models of its class. The authors argue that this efficiency-performance combination is enabled by: (1) a depth-scaled residual connection (DSRes) that preserves activation norms across depth, (2) an adaptive expert bias update schedule that improves routing stability at scale, (3) domain-specific data curation pipelines that preserve high-value technical content, (4) a two-stage midtraining pipeline that extends context to 1M tokens without degrading general capabilities, and (5) a large-scale asynchronous RL campaign that shows continued improvement without plateau. The paper also claims that Beam achieves the best safety-helpfulness tradeoff among open-weight models on the Pareto frontier.

The premise it rests on. The claim that Beam's RL training shows 'no sign of a plateau' and that capabilities 'continued to improve as we increased RL compute' is based on a single large RL run, and the paper's own Figure 4 shows that some benchmarks plateau or decline in later stages, suggesting this is an overgeneralization.

If the paper is right (Pith's inferences, not the authors')

If Beam's scaling claims hold, it suggests that the combination of DSRes, adaptive expert bias updates, and quality-aware data curation can produce frontier-level reasoning in a model with only 23.4B active parameters, potentially democratizing access to advanced AI capabilities.

The reported 3-4x inference efficiency gain over comparable models would make Beam significantly more cost-effective for deployment in coding, mathematics, and agentic applications, potentially shifting the cost-performance frontier for open-weight models.

The claim that RL capabilities continued to improve without plateau over 100M rollouts suggests that further scaling of RL compute could yield additional gains, though the paper acknowledges this is based on a single run.

The safety results (88.9% harmful refusal, 81.5% harmless compliance) suggest that strong safety alignment can be achieved alongside frontier-level capabilities in open-weight models, potentially influencing safety practices in the open-source community.

The paper's claim that RL shows 'no sign of a plateau' is based on a single large run; the authors acknowledge that some benchmarks (e.g., BrowseComp) plateau or decline in later stages. A reader should treat the 'no plateau' claim as a description of the observed trend rather than a proven law.

The depth-scaled residual connection (DSRes) is presented as a drop-in replacement for standard residual connections in pre-normalization architectures. If validated, this technique could be adopted broadly in large-scale transformer training to improve stability without architectural changes.

The paper's data curation methodology—particularly domain partitioning and quality-aware upsampling—could be applied independently of the model architecture to improve training data quality for other large language models.

The multi-teacher on-policy distillation (MOPD) approach for combining agentic and alignment capabilities into a single model could serve as a template for other organizations seeking to merge specialized expert models.

Check this review

What was read

The passes were given 599,600 of the 609,632 characters of the paper's LaTeX source.

Fingerprint of that text (SHA-256): df087627d6664a0cb24150e999823cf601ccf43419608451c7306b747f4b9a72

Pith does not keep that text. Rebuilt from the paper's source or PDF by the same code, it has this fingerprint if it is the same text. From a copy of Pith's code:

PYTHONPATH=. python scripts/review_input_check.py 2611.26358

The passes, in the order they ran

  1. Reader deepseek-v4-flash · medium effort · $0.015 · instructions of the Reader, fingerprint f9ead7300382…
  2. Skeptic deepseek-v4-flash · medium effort · $0.023 · instructions of the Skeptic, fingerprint 8f88c0f61488…
  3. Referee deepseek-v4-flash · medium effort · $0.015 · instructions of the Referee, fingerprint a852d99aea0d…
  4. Check against the paper deepseek-v4-flash · medium effort · $0.018 · instructions of the Check against the paper, fingerprint 6c7a6883d551…
  5. Circularity check deepseek-v4-flash · medium effort · $0.017 · instructions of the Circularity check, fingerprint e9946e39d67b…
  6. Summary for the page deepseek-v4-flash · medium effort · temperature 0.2 · $0.014 · instructions of the Summary for the page, fingerprint e4f56bf568e7…

The passes cost $0.102 in model calls together.

Version stamped on this review: pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.ce10e2b95288

Each link opens a pass's instructions as plain text, the exact bytes its fingerprint was taken over. Every stored set of instructions.

curl -s https://pith.science/methodology/prompts/f9ead730038216984cfb56d7cb34332bdf691828296a1de4245f3d6667d4bf56 | sha256sum

prints the fingerprint of the reader's instructions. The same works for every pass.

The receipt

The review's fingerprint was taken on 20 Nov 2026. The review you are reading still matches it.

Fingerprint of the review (SHA-256): a85f6b0f88af8fc4ffc61a200cd4b8cedd6ddf293c562ce195e9a1f099384b16

curl -s 'https://pith.science/api/v1/verdicts/2611.26358?form=canonical' | sha256sum

prints the fingerprint of the review as it stands now. The rule, in four steps.

A matching fingerprint shows the review has not changed since that date. It does not show that the review is right.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, from the paper's LaTeX source.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper · Review history (2 versions)

The paper

Abstract

Beam is a sparse Mixture-of-Experts (MoE) model with 501.3B total parameters and 23.4B active parameters, specifically built and optimized for coding, reasoning, and agentic workloads. Beam delivers frontier level reasoning, while requiring 3-4x less inference compute than many models of its class. Pretrained on 23.8T diverse, curated, high-quality tokens, Beam features a 1M-token context window achieved through a novel two-stage midtraining phase, built on verifiable, correctness-anchored data. Post-training leveraged a large-scale, fully asynchronous reinforcement learning (RL) campaign across 10.5K GB300 GPUs, generating over 100M rollouts over 4 weeks of training. We believe this is one of the largest scale RL runs conducted by any open lab to date. Across our evaluation suite, capabilities continued to improve as we increased RL compute, with no sign of a plateau. Supported by deep hardware-model co-design, FP8/NVFP4 quantization, and a 1.93B-parameter speculative decoding drafter (DFlash 2), Beam maximizes serving throughput with efficient, flexible deployment. Averaged across conversation turns and harm categories, Beam refuses or pivots on 88.9% of harmful requests while complying with 81.5% of harmless requests, achieving the highest refusal/pivot rate on harmful requests among open-weight models on the Pareto frontier.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-20

The authors' main claim, as the classifier picked it out 0.53

“Beam delivers frontier level reasoning, while requiring 3-4x less inference compute than many models of its class.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.12 · states limitation 0.11 · human participants 0.06 · controlled comparison 0.15 · randomized assignment 0.02 · states quantitative result 0.98 · compares with prior results 0.82 · mentions code or data availability 0.12

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Beam is a sparse Mixture-of-Experts (MoE) model with 501.3B total parameters and 23.4B active parameters, specifically built and optimized for coding, reasoning, and agentic workloads. method 0.59 0.80 0.07 asserted 1.00
Beam delivers frontier level reasoning, while requiring 3-4x less inference compute than many models of its class. result 0.93 0.76 0.25 asserted 1.00
Pretrained on 23.8T diverse, curated, high-quality tokens, Beam features a 1M-token context window achieved through a novel two-stage midtraining phase, built on verifiable, correctness-anchored data. method 0.92 0.91 0.08 asserted 1.00
Post-training leveraged a large-scale, fully asynchronous reinforcement learning (RL) campaign across 10.5K GB300 GPUs, generating over 100M rollouts over 4 weeks of training. method 0.98 0.95 0.05 asserted 1.00
We believe this is one of the largest scale RL runs conducted by any open lab to date. interpretation 0.82 0.69 0.06 hedged possibility 0.57
Across our evaluation suite, capabilities continued to improve as we increased RL compute, with no sign of a plateau. result 0.93 0.95 0.19 asserted 1.00
Supported by deep hardware-model co-design, FP8/NVFP4 quantization, and a 1.93B-parameter speculative decoding drafter (DFlash 2), Beam maximizes serving throughput with efficient, flexible deployment. method 0.91 0.79 0.05 asserted 1.00
Averaged across conversation turns and harm categories, Beam refuses or pivots on 88.9% of harmful requests while complying with 81.5% of harmless requests, achieving the highest refusal/pivot rate on harmful requests among open-weight models on the Pareto frontier. result 1.00 0.93 0.20 asserted 1.00

What kind of paper the classifier reports this is: method or tool 0.96 · evidence from observation or experiment 0.93

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-20: 1 request, 6,478 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (72)
Figure 1 of the paper
Figure 1. Beam’s expert utilization over pretraining. The curve shows the busiest expert’s load divided by the uniform-routing load, averaged across MoE layers; the dashed line marks perfect balance. The ratio reaches 1.04× at pretraining completion. • Scale-Invariant Drop-In Replacement: Because RMSNorm is approximately scaling invariant (i.e., RMSNorm(𝐱 ⋅ 𝑐) ≈ RMSNorm(𝐱)), DSRes preserves the input distribution or the subsequent sublayer outputs in pre-normalization models. • Efficient Recurrent Formulation: DSRes can be computed in a recurrent way: 𝐱𝑛 = √ 𝑛 − 1 √ 𝑛 𝐱𝑛−1 + 1 √ 𝑛 𝐡𝑛 (2
Figure 2 of the paper
Figure 2. Beam training loss in bits per byte (logarithmic scale) over the course of its pretraining run
Figure 3 of the paper
Figure 3. Pretraining performance across scales on decontaminated code and web validation sets. Beam’s recipe scales predictably across four orders of magnitude in compute and achieves Pareto-optimal validation loss versus training compute among the comparable open base models shown. Validation loss is measured in bits per byte; lower is better, same loss at lower FLOPs is also better. Beam Base is shown with a larger circle. Data category Unique tokens (T) Training tokens (T) Mixture share (%) Avg. epochs English web 5.519 8.331 35.00 1.51× Multilingual web 1.669 1.191 5.00 0.71× Code 3.647…
Figure 4 of the paper
Figure 4. Web data curation from Common Crawl and third-party crawls through domain-specific filtering and quality partitioning. Technical domains are partitioned early and treated separately thereafter. using fastText’s lid-176 language identifier. Each partition was then fuzzy-deduplicated against itself using the deduplication pipeline described in Section 2.2.4
Figure 5 of the paper
Figure 5. Data curation as a quality-volume tradeoff. The x-axis shows cumulative tokens removed, and the y-axis shows the token-weighted mean judge quality of the retained data. Each curve progressively removes documents in ascending order of an individual quality signal, while markers show complete multi-rule filter pipelines. Random removal preserves mean quality in expectation (gray dashed line), and removal in ascending judge-quality order defines the oracle upper envelope. 4. Ordinal Evaluation: A pairwise-comparison tournament, run by an internally configured “oracle” judge model,…
Figure 6 of the paper
Figure 6. Compute efficiency of English web data ablations that vary the quality classifier used to assign documents to quality tiers. Classifier arms use a common quality upsampling schedule; DCLM fastText defines the 1.0 baseline. A uniform-sampling control over the retained quality tiers from our classifier isolates the schedule’s contribution. STEM and code validation serve as expected-neutral controls because only English web data differs. fixed and optimize the upsampling schedule. Using our final quality classifier, we partitioned the corpus into quality tiers and study both choices…
Figure 7 of the paper
Figure 7. illustrates the pipeline: a controller plus a fleet of mini-clusters. The controller is a single Ray/KubeRay process per run that owns the workset, decides how work is partitioned, launches and monitors mini-clusters, and reconciles what finished against what was assigned. Each mini-cluster is a self-contained Ray cluster that processes one slice of the corpus to completion and is torn down. Colocating CPU and GPU Workloads. A mini-cluster maps to a given GPU node (H100, H200, RTX Pro 6000, etc). Both rendering and OCR run on the same nodes: the render workers (CPU actors) are…
Figure 8 of the paper
Figure 8. Example of structured OCR output. The input page and detected regions are shown alongside the rendered HTML, preserving text, tables, and equations. • Defective Documents: Documents that are deemed unsuitable (corrupt, encrypted, or over the page limit) go to a dead-letter table with a specified reason listed. As a part of failure handling, we keep a thorough record of each action. Each document writes status events to a status tracking table with the attempt number, final status, and error code. Each page row records whether OCR succeeded, the error code if it did not, and how long…
Figure 9 of the paper
Figure 9. shows all the quality checks and processes that we have once OCR is done. The goal of post-processing is to ensure that low quality and malformed data does not enter our pretraining corpus. Quality issues can arise from different sources: low quality source, OCR processing artifacts. Low Quality Source. The PDF itself could be relatively low signal for a text model. Content targeted at children with a large number of pictures and limited vocabulary is one example. To address this, we employed custom filters for a minimum number of words per page. 18
Figure 10 of the paper
Figure 10. Schematic examples of (A) exact text repetition, (B) runaway counting, and (C) legitimate repeated structure. These illustrate feature families, not production pages or measured classifier decisions. HTML region labels Dominant region type by character count Page Markdown Extract text signals Core + configured extensions Section-specific model Standardize → logistic score Compare with section threshold Extreme-compression rule Counting-loop rule when configured Flag repetition if any rule fires Nine core numeric signals feed the evaluated model. Extensions depend on configuration
Figure 11 of the paper
Figure 11. Repetition detection combines text-derived numeric features with the dominant HTML region type. A section-specific logistic model and configured rules determine whether to flag a page. Judged against an evaluation set sampled from our data, the detector caught 99.3% of severe repetition cases. We achieved 89.5% precision and 92.0% recall overall, while falsely flagging only 0.06% of clean pages
Figure 12 of the paper
Figure 12. This figure demonstrates an overview of the filter-best-fit algorithm. Oversized documents are first filtered out, then a group of bins is allocated for packing. Each sample is assigned to the bin with the tightest fit. If no fit is found, the group is flushed and a new set of empty bins are allocated. The process continues until all data is exhausted. We adopted Unicron (Section 6) with FlashAttention-4 [40] with zigzag-style sequence reordering for efficient (long-context) training. When used with context parallelism, our attention kernel enforces certain constraints on the…
Figure 13 of the paper
Figure 13. This figure demonstrates the midtraining data curation pipeline. The seed pool feeds preparation and generation stages. Candidate examples are assessed using source references and independent checks, then accepted, discarded, or returned for refinement. This shared generate-assess-refine loop supports task synthesis, trace generation and rewriting, and instructional content generation. corrupted content, malformed formatting, excessive repetition, and preprocessing errors across a wide variety of our data sources and domains. 3.5 Stage 1 Data The final Stage 1 midtraining mixture…
Figure 14 of the paper
Figure 14. The final Stage 2 mixture balances improvements in long-context performance against regressions in other capabilities. The figure shows score differences relative to the Stage 1 checkpoint. carefully selected following natural programming language distribution. We pack the coding files in a repository into a single sequence based on their cross-file dependencies. • Synthetic long-context data. We use a mix of in-house and open-source datasets to enhance the model’s long-range retrieval and reasoning capability. The in-house portion is question-answer (QA) pairs gen￾erated from…
Figure 15 of the paper
Figure 15. Optimization diagnostics during RL, including global gradient norm, per-token KL divergence between the sampling and training policies, and sampled-token negative log-likelihood. Shaded regions indicate training restricted to mathematics, STEM, and competitive programming (light), or software engineering (dark). weaknesses in the evaluation harness. We therefore apply a judge2 with a reward-hacking rubric to rollouts that pass their task verifier. Only the final patch was passed to the judge for evaluation. We additionally monitor the chain-of-thought of the model to detect hacking…
Figure 16 of the paper
Figure 16. DeepSWE pass@1 versus requested reasoning effort (left) and realized output length (right), for the SFT checkpoint and three RL checkpoints. RL improves the trade-off between performance and token use and extends the range of reasoning budgets over which performance continues to improve. The checkpoints at 39% and 79% of the run precede the introduction of a more difficult software-engineering task set at 83%. 76 78 80 82 pass@1 (%) STEM SWE SWE-bench Verified 55.0 57.5 60.0 62.5 SWE-bench Pro 20 25 30 35 40 DeepSWE 50.0 52.5 55.0 57.5 60.0 BrowseComp 0 25 50 75 100 run completion…
Figure 17 of the paper
Figure 17. Held-out evaluation scores during RL (pass@1, reasoning effort 512). Shaded regions mark phases that trained on a single domain: mathematics, STEM and code only (light), and software engineering only (dark). BrowseComp measurements begin at 27% of the run, after a change of search backend. Generalization. We evaluated checkpoints during the run on a fixed suite of benchmarks whose tasks are excluded from the RL data. Each benchmark used the same settings throughout: temperature 1, top-𝑝 0.99, 𝑒 = 512 and a 256K output limit (
Figure 18 of the paper
Figure 18. Semantic Structure of the SWE Task Corpus.
Figure 19 of the paper
Figure 19. Intent distribution across terminal use tasks. Circles denote synthetic tasks; squares denote vendor tasks. • Expert human annotators were used to measure the efficiency of the automated checks and surface misalignment between the intent of the check and the model-produced judgments. • Newly discovered quality issues or systemic flaws observed in batches of vendor or synthetic tasks resulted in definitions being expanded. For instance, we found that a subset of synthetic tasks required the agent to interact with “black-box” binaries that the agent was freely able to access and…
Figure 20 of the paper
Figure 20. Synthetic pipeline for terminal. Verifier Robustness. Verifiers are central to RL: they evaluate an agent’s work and produce the reward used for training. Their quality directly shapes the training signal. For terminal tasks, verifiers were typically custom shell or Python scripts that check an agent’s solution (e.g., inspecting files, validating computed values, or measuring runtime). Good verifiers have three essential properties: alignment with the task description, comprehensive coverage, and robustness to reward hacking. Alignment with the Task Description. The task’s state…
Figure 21 of the paper
Figure 21. Terminal task descriptions generated with low and high task specificity. Task Descriptions. The second stage used a judge to convert seeds into full task descriptions that specifically targeted the contents of the seeds. To control the diversity of the task descriptions, we relied on different “diversity knobs” for generation, including task specificity, and instruction style/tone. These knobs were used to steer the model generation in certain directions and proved useful in preventing diversity collapse. For instance, we could steer the task description toward a technical,…
Figure 22 of the paper
Figure 22. The synthetic task generation pipeline for University Math. courses). For each planned combination, the model was asked to write one question built around exactly those concepts, subject to a set of constraints: the question had to be single-part and self-contained, had to have a numeric or closed-form answer that can be checked automatically, and had to be difficult because of the depth of reasoning required (rather than because of heavy notation or large numbers). The model also needed to solve its own question and revise it if a one- or two-line shortcut existed, and it had to…
Figure 23 of the paper
Figure 23. Distribution of embeddings of the University Math questions after t-SNE. Verification extracted the final submitted program and ran it in an isolated grading sandbox with time limits. The task reward was binary: 1 only if every test passes, otherwise 0. Compilation errors, runtime failures, wrong outputs, and timeouts received no partial credit. Quality and Difficulty. For quality screening, we used deterministic checks, judges,classifiers, reference execution, and deduplication. We checked that questions had a minimum number of test cases, that output comparison scripts aligned…
Figure 24 of the paper
Figure 24. Agentic knowledge-work environments and task curation. The largest sectors in the final pool were finance & insurance, health care & social assistance, and professional, scientific & technical services. Required deliverables included .docx, .xlsx, .md, .csv, .pdf, and .pptx, alongside other formats. Many tasks required multiple output formats, and tasks were evaluated against weighted rubric criteria, including penalty criteria for undesirable outputs. Every task was authored as a Harbor task directory and ships as a digest-pinned container image built from its environment,…
Figure 25 of the paper
Figure 25. Predictable RL Evolution. Left: Predicted reward gains from the SFT checkpoint, compared to realized gains through RL, across components targeting different principles. Predictions are made using a replicator equation with a learning constant fit to a single run. Scatter shows distribution across 318 post-training runs, large symbols represent the final alignment teacher. Right: Same equation applied to forecasting grader-detected misbehaviors, shaded band = forecast ± SE, dots = observed. Changes in reward function (green circles) accurately predict squashing of misbehavior.…
Figure 26 of the paper
Figure 26. Knowledge Calibration Training Reduces Hallucinations and Increases Honest Abstentions.
Figure 27 of the paper
Figure 27. How one SQL-grounded RL item is produced
Figure 28 of the paper
Figure 28. LongBench-v2 accuracy versus total parameter count (log scale). † indicates externally reported results; all others were reproduced internally. Beam is served in NVFP4 with an FP8 KV cache. response satisfies it. The constraints cover response length; output format; required or forbidden vocabulary; per-sentence or per-section templates; capitalization and punctuation; and ordering or repetition patterns. 56
Figure 29 of the paper
Figure 29. RLAIF Pushes the Open Pareto Frontier on Conversational Quality
Figure 30 of the paper
Figure 30. Our multi-teacher on-policy distillation (MOPD) setup. the contexts it visits, with the supervising teacher selected according to the task family. This approach allows each expert to be trained and evaluated independently before their capabilities are consolidated into the final model. 4.3.1 Defining the Distillation Objective For a batch of 𝑀 tasks 𝑥𝑖 ∼ 𝒟, the student samples 𝑅 rollouts 𝑦𝑖𝑗 ∼ 𝜋𝜃 (⋅ ∣ 𝑥𝑖 ) per task. Each rollout is scored by one selected teacher 𝜋𝑇𝑖𝑗 . With 𝑅 rollouts per task, we average the full-vocabulary reverse KL over trainable tokens across the batch: ℒˆ(𝜃) =
Figure 31 of the paper
Figure 31. Blockwise reverse KL processes paired 𝐶 × 𝑉 student and teacher logit chunks, reducing temporary logit memory from 𝑂(𝑁𝑉) to 𝑂(𝐶𝑉). What Student Probabilities to Use The student that generates a rollout can differ from the student being updated: while inference is running, the trainer continues to learn from other rollouts. We must therefore decide which version’s disagreement with the teacher the loss should measure. We use the trainer’s current probabilities for both the student log probabilities and the weighting over vocabulary entries in the KL. Consider a context where the…
Figure 32 of the paper
Figure 32. shows the main components and their interactions. The control service manages the job lifecycle and coordinates rollout generation, batch assembly, and policy updates. Rollout workers execute agent–environment interactions using policy inference and sandboxes for tool execution, with a separate judge model reviewing solutions where enabled. The training service consumes completed batches and updates the policy, while the control service coordinates the publication and adoption of new policy versions. Storage preserves checkpoints and execution records. Each job runs in a dedicated…
Figure 33 of the paper
Figure 33. Policy updates during a multi-turn rollout. Within a policy version, subsequent turns reuse cached context and prefill only newly added tokens. When weights change during decoding, the running turn retains its existing KV cache. At the next turn, the full context is recomputed under the updated weights using previously recorded expert selections. Widths are schematic. Weight Updates and Prefill Expert Replay. Long rollouts can span multiple policy updates (
Figure 34 of the paper
Figure 34. Overlap between rollout generation, batch preparation, and training. The next batch is prepared while the trainer processes the current batch, and updated policy weights become available as training progresses. Inference workers adopt updates asynchronously. The one-batch lookahead is illustrative and does not bound policy age to one version. Widths are schematic. Weight Distribution. Frequent policy updates must be transferred from the trainer to inference workers that may use a different sharding layout. We transfer weights directly from trainer memory using NIXL [116], avoiding…
Figure 35 of the paper
Figure 35. Sources of policy staleness across production jobs. The stacked components measure policy-version differences within rollouts, between rollouts in a group, between groups for a task, and the remaining gap at training consumption. Differences between rollouts in the same group were the largest component. All components share a common normalization: the mean total policy age over valid measurement windows is set to 1. The horizontal axis concatenates recorded 30-minute windows across twelve jobs, omitting gaps between jobs; alternating shaded bands distinguish jobs. The breakdown was…
Figure 36 of the paper
Figure 36. FP8/NVFP4 vs. BF16 Beam 1.0 benchmark differences. Each pair of bars compares FP8 (upper) and NVFP4 (lower) with BF16 on one of 22 benchmarks across coding, reasoning, search/tools, and general capabilities. Bar length shows the score difference (quantized − BF16) in percentage points: positive values indicate higher scores than BF16, and negative values indicate lower scores. A label of 0.0 indicates a match to one decimal place. Across all 22 benchmarks, the mean difference is +0.03 points for FP8 (higher on 9, lower on 11) and −0.45 points for NVFP4 (higher on 4, lower on 16).…
Figure 37 of the paper
Figure 37. Memory allocation per B300 GPU (268.6 GiB) after model load. Each configuration reserves 90% of HBM; KV cache uses the space remaining after weights and other allocations. Labels inside bars show GiB; right-hand labels show replica-wide KV capacity in full-attention tokens. Rows are sorted by KV capacity
Figure 38 of the paper
Figure 38. 8k/1k scenario: Output throughput per GPU versus interactivity on NVIDIA B300 GPUs. Each point represents a concurrency level, labeled c8–c256 by the number of concurrent requests. Panel (a) uses each checkpoint’s smallest deployment: four GPUs for FP8 and NVFP4, and eight for BF16. Panel (b) uses eight GPUs for all checkpoints. Dotted lines mark 50 and 100 tok/s/user. Points higher and farther right indicate better performance. NVFP4 provides the highest per-user speed at eight concurrent requests. It reaches 199 tok/s/user on eight GPUs and 168 on four GPUs, compared with 126 for…
Figure 39 of the paper
Figure 39. AgentX replay: output throughput per GPU versus interactivity on NVIDIA B300 GPUs. Points are labeled by the number of concurrent sessions. Panel (a) uses four GPUs for FP8 and NVFP4, and eight for BF16; panel (b) uses eight GPUs for all checkpoints. The dotted line marks 50 tok/s/user. Points higher and farther right indicate better performance. FP8 also improves throughput at 50 tok/s/user. It delivers 1.7× BF16’s per-GPU throughput on four GPUs and 1.5× on eight GPUs. Across all configurations, median time to first token at 64 sessions ranges from 189 to 224 ms. Deployment…
Figure 40 of the paper
Figure 40. Enhanced token efficiency through RL. DeepSWE pass@1 against mean output and tool tokens per episode; each curve is one RL checkpoint (darker is later), each point a reasoning effort from 64 to 512. engineering [138, 139]. We separately evaluate on the original 731-task public test split (v1), and on the 51- task HARD subset of its revised public release (v2). The HARD subset contains tasks that at least two of five reference model families failed under Scale’s evaluation protocol, after ambiguous tasks were excluded. SWE Atlas Codebase QnA evaluates code comprehension through…
Figure 41 of the paper
Figure 41. This plot shows a comparison of Beam Base against DeepSeek V4 Pro Base, Nemotron 3 Ultra Base, MiMo-V2.5-Pro Base for long context capability. Beam outperforms the baselines up to the trained 512K context length, but also generalizes well to 1M context as measured with RULER. Beam DeepSeek V4 Pro Nemotron 3 Ultra MiMo-V2.5-Pro higher = better Score 0% 20% 40% 60% 80% 100% Logical understanding 82.4 78.7 80.6 78.4 Math & STEM 61.4 49.3 55.6 47.9 Coding 68.3 64.4 57.8 65.8 SWE 84.2 81.2 81.6 81.0 Terminal 69.2 56.0 57.9 56.6
Figure 42 of the paper
Figure 42. , the gap between Beam and other base models is more prominent in the held-out in-house evaluations, outperforming other base models, even the larger MiMo and DeepSeek V4 Pro baselines across the board. 6 Infrastructure Approach 6.1 Introduction Training Beam required infrastructure that could sustain distributed model computation while supplying data, serving policy inference, and executing agent tasks at scale. The workload changed across training stages: pretrain￾ing consumed materialized sequences, while RL coupled training with long-running rollouts, tool execution, and reward…
Figure 43 of the paper
Figure 43. Shared training machinery and stage-specific execution paths. Pretraining and midtraining share a supervised training path; RL and on-policy distillation share asynchronous rollout infrastructure with different training objectives. Judge services support selected RL tasks, while teacher annotation supplies the distillation targets. Both paths supply microbatches and objectives to a common Unicron engine. Supporting infrastructure spans both paths. The diagram represents separate training configurations; policy-weight publication is specific to the online training path and is…
Figure 44 of the paper
Figure 44. Example of productive-time and interruption accounting during the 21.9-day pretraining run on 6,144 GPUs. The timeline illustrates how retained progress, recovery, and operational events contribute to the goodput measurement. fault that took down several NVL72 racks during one high-compute agentic RL run lost ∼7% of the groups in a batch. We considered this loss acceptable given the size of the fault domain. Tracking Interruptions and Retained Progress
Figure 45 of the paper
Figure 45. Sandbox architecture. The stateless gateway places new sandboxes in an eligible cluster using routing policy and advertised capacity; its response identifies the sandbox and owning endpoint. Each sandbox-server replica implements both lifecycle control and command proxying. Pod and Node watches feed the capacity snapshots read by the gateway. Commands and deletion address the owning cluster directly. Each sandbox contains one task container with the daemon and task processes, isolated by gVisor or Kata Containers with Cloud Hypervisor (CH). Shared Layers and Prewarming. We organize…
Figure 46 of the paper
Figure 46. groups the failure categories. Recovery Outcomes. Inference failures are handled by rescheduling affected serving pods on healthy nodes (Section 6.7.4). Trainer failures require a job-level restart (Section 6.7.6). Some inference failures required manual intervention to stop a repeatedly crashing SGLang serving instance or mark a node unhealthy when automation had not done so in time. Of 99 recorded failures in the high-compute agentic RL deployment, 89 (89.9%) were resolved automatically. In two cases (2.0%), automatic recovery failed and manual intervention restored service.…
Figure 47 of the paper
Figure 47. Recovery intervals normalized independently to 100%. Left: a local serving-instance repair ending at an ordinary completed request. Right: a coordinated job restart ending at its first completed trainer step. Phase widths are relative to each case’s own interval; the two bars do not compare absolute duration or imply representative recovery percentiles. whether recovery was automatic or manual. These percentages describe recovery outcomes for the recorded failures; they do not measure availability or failure frequency. The scope of recovery determines what must resume: local…
Figure 48 of the paper
Figure 48. Safety SFT dataset composition. The reward comes from a context-aware judge of refusal behavior that is instantiated differently per task type: on multi-turn tasks, it sees the prior conversation; on jailbreaks, the seed intent behind the attack; and on agentic tasks, the prior turns plus the unsafe action. Tasks the SFT checkpoint already handles reliably are dropped from the mixture, so the run spends its budget on cases that are still open. The Beam safety RL data and reward details are described below. Data. The Beam safety RL data is structured along two axes: the…
Figure 49 of the paper
Figure 49. Task category distribution in Beam safety RL data. Judge and Reward. The Beam safety RL judge is context-aware and is designed to score model responses whether the task is harmful or harmless. This is similar to the setup in Iris Safety Evaluations (ISE) (
Figure 50 of the paper
Figure 50. Molecular Biology Capabilities Test (MBCT). Virology and Pathogen-Related Knowledge Virology Capabilities Test (VCT). The Virology Capabilities Test evaluates laboratory methods and experimen￾tal troubleshooting with potential dual-use relevance while excluding explicitly hazardous material. Accordingly, it measures risk-relevant virology knowledge rather than end-to-end biological misuse capability. Interpretation and Limitations. VCT measures performance on practical virology knowledge. It does not directly assess willingness to provide harmful assistance, successful laboratory…
Figure 51 of the paper
Figure 51. World-Class Biology (WCB). Interpretation and Limitations. The Human Pathogen Capabilities Test (HPCT) assesses a model’s knowledge of laboratory methods and experimental troubleshooting involving human pathogens. Accuracy excluding refusals should be interpreted alongside raw accuracy and refusal rates. The human reference accuracy noted below likewise reflects performance on expert-specific question subsets rather than the complete evaluation set. Dataset and Scoring. We evaluate all 100 text-only questions in the July 9, 2026 release. The questions assess practical knowledge of…
Figure 52 of the paper
Figure 52. Virology Capabilities Test (VCT-V2). We report their results differently, because they assess different stages of this process and use different scoring methods. ABC-Bench Fragment Design. We assess DNA design capabilities using the Fragment Design task from SecureBio’s Agentic Bio-Capabilities Benchmark (ABC-Bench). In this task, an agent is provided with a target DNA sequence and must design shorter fragments suitable for commercial synthesis and subsequent reconstruction via Gibson Assembly. While this capability is foundational to legitimate biological research, it also…
Figure 53 of the paper
Figure 53. Human Pathogen Capabilities Test (HPCT). above. Higher scores indicate higher proficiency in autonomously generating procedurally correct, synthesis-ready DNA fragment designs. Findings. As shown in
Figure 54 of the paper
Figure 54. ABC-Bench Fragment Design. Interpretation and Limitations. These results characterize model performance against the specific, localized screening tools and reference databases available to the agent during the evaluation. Because no physical DNA is ordered or experimentally synthesized, the scores do not guarantee that commercial synthesis providers would accept the orders, nor do they prove that the reconstruction plan would succeed in an actual laboratory environ￾ment. Furthermore, the limited dataset size (𝑁 = 10) and the reliance on LLM-based evaluation constrain the…
Figure 55 of the paper
Figure 55. ABC-Bench Screening. To assess broader behavioral differences, we reviewed models’ tool use, environment inspection, and reasoning about the evaluator. Most models relied primarily on local screening; some researched assembly methods online, while Beam, Inkling 1T, and DeepSeek V4.1 Flash searched the sandbox for task files or information about the evaluator. None of these searches uncovered hidden grading materials or other unintended information. These behaviors also did not consistently track jury scores. Scores instead reflected the submitted design strategy: DeepSeek V4 Pro…
Figure 56 of the paper
Figure 56. Biological Targeted Information for Exclusion and Refusal Benchmark (BioTIER). framework to ensure that the final rubric itself would not constitute a potential source of information hazard. The experts weighed the relative importance of each indicator to the overall success of a protocol, and a scoring system involving individual and sequential values was included to aid the protocol evaluation. Ultimately, the weighting criteria can be adjusted by the client depending on their relative interest in individual vs. sequential success in meeting the indicators. Participants had five…
Figure 57 of the paper
Figure 57. Successful refusal rate by target set. A subject-matter expert applied the rubric created by Nemesys Insights to each participant’s final protocol. We compared composite scores between the Beam and Kimi K3 Max Reasoning groups to assess relative rubric-scored assistance under the conditions tested. Findings. Participants using Kimi K3 Max Reasoning achieved higher scores than participants using Beam on the reconstruction task, with a mean final score of 6.6 compared with 0.4. The observed mean difference was 6.2 points (Figures 58 and 59). However, the difference was concentrated…
Figure 58 of the paper
Figure 58. Uplift study—total score by participant
Figure 59 of the paper
Figure 59. Uplift study—mean participant score. and the uplift trial and relevant third-party assurance evaluation showed no gain in scale, sophistication, and accessibility of harmful activity over a comparable open model. Consistent with the marginal-risk approach, this determination weighs these results alongside broader societal factors, including the real-world barriers to misuse that lie outside the model and the wide existing availability of comparable capability. We therefore assess Biological Weapons as Limited under the OSF. 7.3.2 Chemical Weapons WMDP-Chem We assess…
Figure 60 of the paper
Figure 60. Weapons of Mass Destruction Proxy (chemistry). 7.3.3 Loss of Control We have evaluated the Loss of Control risk as Limited under the OSF on the basis that there is no material erosion of the following three properties: • Monitorability: The ability to audit what a model has done. • Evaluability: The ability to measure model behaviors before release. • Controllability: The ability to intervene on a model’s behavior. Note that under the OSF, Loss of Control risk may be Critical if a model resists correction beyond the operator’s control, such that operators can no longer dependably…
Figure 61 of the paper
Figure 61. Monitorability results for Beam and GLM-5.3-Flash. We use a fixed frontier model as a monitor for all tasks and models. model when completing a task (like the chain of lemmas used to prove a theorem) and intervention monitorability is the ability to identify properties of a model’s behavior, such as if it is fabricating an answer or not. We find it useful to further categorize intervention monitorability by the property that is being monitored: • Task gaming is when a model uses a planted answer as a shortcut to solving a problem. Planted answers can look like leaked grader…
Figure 62 of the paper
Figure 62. Verbalized evaluation awareness rates across our Iris Safety Evaluations (single-turn, multi-turn, and agentic) as well as our deployment evaluation based on WildChat. Verbalized evaluation awareness rates are calculated using a GLM-5.3-Flash based monitor over the model’s reasoning and final response. • Evaluation tampering (4 scenarios): Rigging an assessment of itself that decides whether it stays de￾ployed (e.g. by falsely reporting completion, reading a protected answer key, dropping hard test cases or suppressing a reviewer’s adverse finding) instead of accepting an honest…
Figure 63 of the paper
Figure 63. Loss of Control automated audit results. Each model is evaluated by an automated auditor driven by a frontier model twenty times over several scenarios for each behavior and then graded on a scale between 1–10 where a 10 means the model carried out the misaligned behavior in full. This plot shows the elicitation rate, or the percent of audits scoring 7 or above. Error bars show 95% bootstrap confidence intervals. Interpretation and Limitations. These audits measure propensity rather than capability: how often a model takes an available opportunity to weaken a control over itself in…
Figure 64 of the paper
Figure 64. CTF challenge solve rate breakdown by challenge difficulty, challenge group, and challenge type. Beam outperforms all Western open-source models, but does not reach the cyber offense capabilities of the leading open-source models. 7.4 Iris Safety Evaluations Iris Safety Evaluations (ISE) is Reflection’s in-house, soon-to-be open-sourced safety evaluation suite, combining single-turn, multi-turn, jailbreak, and agentic evaluations to measure adherence to our safety policies. To assess how deployment settings affect safety behavior, ISE supports (1) base model only mode (2) search…
Figure 65 of the paper
Figure 65. Beam is on the observed safety-helpfulness Pareto frontier. all conversation turns and harm categories. Beam refuses or pivots on 88.9% of harmful requests (C3 or C4) while complying with 81.5% of harmless requests (C1 or C2). Compared to other open-weight models, Beam leads both the Chinese and Western models on the observed safety-helpfulness Pareto frontier, and has the highest refusal and pivot rate on harmful requests. We observe similar results across different reasoning efforts for Beam. We also analyze Beam’s refusal behavior across control layers in
Figure 66 of the paper
Figure 66. Beam’s refusal behavior across harm levels (control layers). By Harm Category: Refusals or Pivots on Harmful Requests (C3 or C4) We further report Beam’s refusal/pivot rate on harmful requests (C3 or C4) across harm categories in Tables 30 to 32. Across every harm category, Beam refuses or pivots on more than 80% of harmful requests. Model H1 H2 H3 H4 H5 Beam (BF16) 90.6% 83.3% 86.2% 87.5% 85.5% Beam (FP8) 90.6% 83.8% 86.2% 88.5% 88.2% Beam (NVFP4) 90.3% 82.1% 87.8% 86.0% 84.7% GPT-OSS-120B 81.7% 78.1% 85.1% 83.6% 88.5% Nemotron 3 Ultra 85.6% 67.1% 79.5% 82.5% 72.2% Inkling 1T…
Figure 67 of the paper
Figure 67. Evaluating ISE harmful requests with search and adversarial instructions. Model Refusal Pivot Clarification Answered Beam 2.3% 1.1% 1.1% 94.7% GPT-OSS-120B 4.3% 0.5% 1.6% 93.6% Nemotron 3 Ultra 2.7% 1.1% 1.5% 94.6% Inkling 1T 2.1% 0.8% 1.0% 96.0% Muse Glimmer 4.2% 1.0% 0.9% 93.8% Kimi K2.6 2.6% 0.8% 0.8% 95.7% Kimi K3 2.4% 1.3% 1.4% 94.7% GLM-5.2 2.2% 0.6% 0.8% 96.4% GLM-5.3 2.2% 1.2% 1.0% 95.6% DeepSeek V4 Flash 2.0% 0.7% 1.0% 96.2% DeepSeek V4 Pro 1.7% 0.6% 1.1% 96.5% DeepSeek V4.1 Flash 1.7% 0.9% 0.9% 96.3% Qwen 3.8 Max 2.3% 0.9% 1.0% 95.7%
Figure 68 of the paper
Figure 68. Each episode pairs a benign user task with an instruction planted in a tool result. We grade whether the user’s task was completed (utility) and whether the injected goal was carried out (attack success rate, ASR). Model MirrorDojo ASR ↓ MirrorDojo utility ↑ clean / attacked MirrorDyn ASR ↓ MirrorDyn utility ↑ clean / attacked Beam 3.1% 87.6 / 86.9 3.6% 83.3 / 82.7 DeepSeek V4.1 Flash 0.0% 91.8 / 88.9 0.0% 81.7 / 78.2 Kimi K3 0.0% 93.8 / 91.9 0.2% 81.7 / 76.2 GLM-5.3 0.1% 93.8 / 90.9 0.0% 85.0 / 80.0 Qwen 3.8 Max 0.6% 92.8 / 91.4 0.4% 81.7 / 81.1 DeepSeek V4 Pro 0.9% 91.8 / 89.3…
Figure 69 of the paper
Figure 69. Gray Swan Shade. All settings Coding agent Tool use 22.1% 18.5% 25.5% 73.1% 65.9% 79.9% 78.4% 71.6% 84.7% Gray Swan IPI Arena lower = more robust · transfer attack success on Q1+Q2 2026 arena scenarios ASR@1 ASR@10 ASR@15
Figure 70 of the paper
Figure 70. Gray Swan IPI Arena and unintended financial transactions. The scenarios are designed to match the difficulty of real tasks frontier models can do today, including coding, computer use, and tool use. 7.7.3 Gray Swan Public IPI Black Challenge Bug Bounty Beam is participating in Gray Swan’s three-week, large-prize public bug bounty on indirect prompt injection (IPI) against production agent harnesses. Reflection is co-sponsoring with other frontier labs to support the safety ecosystem and promote improvements to future models and safeguards. This is the largest and most realistic AI…
Figure 71 of the paper
Figure 71. Gray Swan Reward Hacking Bench. 7.7.5 Irregular Independent Cyber External Evaluation Irregular conducted an external evaluation of Beam’s offensive cybersecurity capabilities using private challenges covering vulnerability research and exploitation, network attacks, evasion, and longer-horizon cyber operations. Across individual-task categories, average success rates ranged from 43% to 60%, while no end-to-end CyScenari￾oBench scenarios were completed. This pattern indicates useful capabilities on bounded tasks alongside clear limits on autonomous, multi-step operations. These…
Figure 72 of the paper
Figure 72. Learning rate and weight decay tuning. (a) Validation loss vs effective learning rate for four models. Curves are quadratic fits to grid of effective learning rate parameterized by 𝛼 where 𝜂 ′ ef f = 𝛼𝜂ef f,base (𝜂ef f,base is starting to have an effective learning rate to tune around). The best effective learning rate (𝜂 ∗ ef f = 𝛼 ∗𝜂ef f,base) is indicated with a vertical dashed line. (b, c) Scaling laws for best peak learning rate 𝜂 ∗ max = 𝛼 ∗𝜂max,base and best weight decay 𝜆 ∗ max = (𝛼 ∗ ) 2𝜆max,base. 𝜂 ′ max = 𝛼𝜂max,base (A.18) Thus, to tune effective learning rate

Figures are the authors', reproduced from the paper.

References (199)
  1. [1] N.Shazeer,A.Mirhoseini,K.Maziarzetal.,“OutrageouslyLargeNeuralNetworks: TheSparsely-Gated Mixture-of-ExpertsLayer,”2017.[Online].Available: https://arxiv.org/abs/1701.06538 on PithPith's reviewarXiv
  2. [2] A.Vaswani,N.Shazeer,N.Parmaretal.,“AttentionIsAllYouNeed,”inAdvancesinNeuralInformation ProcessingSystems,vol.30,2017.[Online].Available: https://proceedings.neurips.cc/paper/2017/hash/3f5 ee243547dee91fbd053c1c4a845aa-Abstract.html on Pith
  3. [3] J.Kim,B.Lee,C.Parketal.,“Peri-LN:RevisitingNormalizationLayerintheTransformerArchitecture,” 2025.[Online].Available: https://arxiv.org/abs/2502.02732 on PithPith's reviewarXiv
  4. [4] Gemma Team, S. E. Abd, V. Aggarwalet al., “Gemma 4 Technical Report,” 2026. [Online]. Available: https://arxiv.org/abs/2607.02770 on PithPith's reviewarXiv
  5. [5] Y.Yin,W.Huang,K.Songetal.,“PanguUltra: PushingtheLimitsofDenseLargeLanguageModelson AscendNPUs,”2025.[Online].Available: https://arxiv.org/abs/2504.07866 on PithPith's reviewarXiv
  6. [6] J.Ainslie,J.Lee-Thorp,M.deJongetal.,“GQA:TrainingGeneralizedMulti-QueryTransformerModels fromMulti-HeadCheckpoints,”2023.[Online].Available: https://arxiv.org/abs/2305.13245 on PithPith's reviewarXiv
  7. [7] M.Dehghani,J.Djolonga,B.Mustafaetal.,“ScalingVisionTransformersto22BillionParameters,”2023. [Online].Available: https://arxiv.org/abs/2302.05442 on PithPith's reviewarXiv
  8. [8] Gemma Team, A. Kamath, J. Ferretet al., “Gemma 3 Technical Report,” 2025. [Online]. Available: https://arxiv.org/abs/2503.19786 on PithPith's reviewarXiv
  9. [9] J. Su, Y. Lu, S. Panet al., “RoFormer: Enhanced Transformer with Rotary Position Embedding,” 2023. [Online].Available: https://arxiv.org/abs/2104.09864 on PithPith's reviewarXiv
  10. [10] N.Shazeer,“GLUVariantsImproveTransformer,”2020.[Online].Available: https://arxiv.org/abs/2002.052 02 on Pith
  11. [11] DeepSeek-AI, A. Liu, B. Fenget al., “DeepSeek-V3 Technical Report,” 2025. [Online]. Available: https://arxiv.org/abs/2412.19437 on PithPith's reviewarXiv
  12. [12] B. Zhang and R. Sennrich, “Root Mean Square Layer Normalization,” 2019. [Online]. Available: https://arxiv.org/abs/1910.07467 on PithPith's reviewarXiv
  13. [13] GemmaTeam,M.Riviere,S.Pathaketal.,“Gemma2: ImprovingOpenLanguageModelsataPractical Size,”2024.[Online].Available: https://arxiv.org/abs/2408.00118 on PithPith's reviewarXiv
  14. [14] A. Kazemnejad, I. Padhi, K. N. Ramamurthyet al., “The Impact of Positional Encoding on Length GeneralizationinTransformers,”2023.[Online].Available: https://arxiv.org/abs/2305.19466 on PithPith's reviewarXiv
  15. [15] Y.Bondarenko,M.Nagel,andT.Blankevoort,“QuantizableTransformers: RemovingOutliersbyHelping AttentionHeadsDoNothing,”2023.[Online].Available: https://arxiv.org/abs/2306.12929 on PithPith's reviewarXiv
  16. [16] Z.Qiu,Z.Wang,B.Zhengetal.,“GatedAttentionforLargeLanguageModels: Non-linearity,Sparsity,and Attention-Sink-Free,”2025.[Online].Available: https://arxiv.org/abs/2505.06708 on PithPith's reviewarXiv
  17. [17] OpenAI,S.Agarwal,L.Ahmadetal.,“gpt-oss-120b&gpt-oss-20bModelCard,”2025.[Online].Available: https://arxiv.org/abs/2508.10925 on PithPith's reviewarXiv
  18. [18] K. M. Nakanishi, “Scalable-Softmax Is Superior for Attention,” 2025. [Online]. Available: https: //arxiv.org/abs/2501.19399 on PithPith's reviewarXiv
  19. [19] L.Wang,H.Gao,C.Zhaoetal.,“Auxiliary-Loss-FreeLoadBalancingStrategyforMixture-of-Experts,”2024. [Online].Available: https://arxiv.org/abs/2408.15664 on PithPith's reviewarXiv
  20. [20] Team OLMo, P. Walsh, L. Soldainiet al., “2 OLMo 2 Furious,” 2025. [Online]. Available: https://arxiv.org/abs/2501.00656 on PithPith's reviewarXiv
show 80 more
  1. [21] V.Singh,L.Krauss,S.Jaghouaretal.,“ArceeTrinityLargeTechnicalReport,”2026.[Online].Available: https://arxiv.org/abs/2602.17004 on PitharXiv
  2. [22] A. Defazio, “Why Gradients Rapidly Increase Near the End of Training,” 2025. [Online]. Available: https://arxiv.org/abs/2506.02285 on PithPith's reviewarXiv
  3. [23] I. Loshchilov and F. Hutter, “Decoupled Weight Decay Regularization,” inInternational Conference on LearningRepresentations,2019.[Online].Available: https://arxiv.org/abs/1711.05101 on PithPith's reviewarXiv
  4. [24] Microsoft AI Team, “MAI-Thinking-1: Building a Hill-Climbing Machine,” 2026. [Online]. Available: https://microsoft.ai/pdf/mai-thinking-1.pdf 139 BeamModelTechnicalReport,ReflectionTeametal on Pith
  5. [25] N.Muennighoff,A.M.Rush,B.Baraketal.,“ScalingData-ConstrainedLanguageModels,”2023.[Online]. Available: https://arxiv.org/abs/2305.16264 on PithPith's reviewarXiv
  6. [26] J.Bjorck,A.Benhaim,V.Chaudharyetal.,“ScalingOptimalLRAcrossTokenHorizons,”inInternational ConferenceonLearningRepresentations,2025.[Online].Available: https://arxiv.org/abs/2409.19913 on PithPith's reviewarXiv
  7. [27] S.Malladi,K.Lyu,A.Panigrahietal.,“OntheSDEsandScalingRulesforAdaptiveGradientAlgorithms,” inAdvancesinNeuralInformationProcessingSystems,vol.35,2022.[Online].Available: https://proceeding s.neurips.cc/paper_files/paper/2022/hash/32ac710102f0620d0f28d5d05a44fe08-Abstract.html on Pith
  8. [28] J.W.Rae,S.Borgeaud,T.Caietal.,“ScalingLanguageModels: Methods,Analysis&InsightsfromTraining Gopher,”2021.[Online].Available: https://arxiv.org/abs/2112.11446v1 on PithPith's reviewarXiv
  9. [29] G.Penedo,Q.Malartic,D.Hesslowetal.,“TheRefinedWebDatasetforFalconLLM:OutperformingCurated CorporawithWebData,andWebDataOnly,”2023.[Online].Available: https://arxiv.org/abs/2306.01116 on PithPith's reviewarXiv
  10. [30] J. Li, A. Fang, G. Smyrniset al., “DataComp-LM: In search of the next generation of training sets for languagemodels,”inAdvancesinNeuralInformationProcessingSystems,vol.37,2024.[Online].Available: https://arxiv.org/abs/2406.11794 on PithPith's reviewarXiv
  11. [31] G.Penedo,H.Kydlíček, L.B.allaletal., “TheFineWebDatasets: DecantingtheWebfortheFinestText DataatScale,”2024.[Online].Available: https://arxiv.org/abs/2406.17557 on PithPith's reviewarXiv
  12. [32] TeamOlmo,A.Ettinger,A.Bertschetal.,“Olmo3,”2025.[Online].Available: https://arxiv.org/abs/2512.1 3961v1 on Pith
  13. [33] A. Joulin, E. Grave, P. Bojanowskiet al., “Bag of Tricks for Efficient Text Classification,” 2016. [Online]. Available: https://arxiv.org/abs/1607.01759 on PithPith's reviewarXiv
  14. [34] A. Joulin, E. Grave, P. Bojanowskiet al., “FastText.zip: Compressing Text Classification Models,” 2016. [Online].Available: https://arxiv.org/abs/1612.03651 on PithPith's reviewarXiv
  15. [35] A. Wettig, K. Lo, S. Minet al., “Organize the Web: Constructing Domains Enhances Pre-Training Data Curation,”2025.[Online].Available: https://arxiv.org/abs/2502.10341 on PithPith's reviewarXiv
  16. [36] A.H.Kargaran,A.Imani,F.Yvonetal.,“GlotLID:Languageidentificationforlow-resourcelanguages,”in FindingsoftheAssociationforComputationalLinguistics: EMNLP2023. AssociationforComputational Linguistics,2023,pp.6155–6218.[Online].Available: https://aclanthology.org/2023.findings-emnlp.410/ on Pith
  17. [37] “go-enry: A Faster File Programming Language Detector,” Software repository. [Online]. Available: https://github.com/go-enry/go-enry on Pith
  18. [38] Y.Fratantonio,L.Invernizzi,L.Farahetal.,“Magika: AI-PoweredContent-TypeDetection,”2024.[Online]. Available: https://arxiv.org/abs/2409.13768 on PithPith's reviewarXiv
  19. [39] “chardetng: A Character Encoding Detector for Legacy Web Content,” Software repository. [Online]. Available: https://github.com/hsivonen/chardetng on Pith
  20. [40] T.Zadouri,M.Hoehnerbach,J.Shahetal.,“FlashAttention-4: AlgorithmandKernelPipeliningCo-Design forAsymmetricHardwareScaling,”2026.[Online].Available: https://arxiv.org/abs/2603.05451 on PitharXiv
  21. [41] D.Narayanan,M.Shoeybi,J.Casperetal.,“EfficientLarge-ScaleLanguageModelTrainingonGPUClusters Using Megatron-LM,” inProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, ser. SC ’21. Association for Computing Machinery, 202... on Pithdoi:10.1145/3458817.3476209
  22. [42] J.Zhang,Y.Bai,X.Lvetal.,“LongCite: EnablingLLMstoGenerateFine-grainedCitationsinLong-context QA,”2024.[Online].Available: https://arxiv.org/abs/2409.02897 on PithPith's reviewarXiv
  23. [43] Y.Bai,X.Lv,J.Zhangetal.,“LongAlign: ARecipeforLongContextAlignmentofLargeLanguageModels,” 2024.[Online].Available: https://arxiv.org/abs/2401.18058 on PithPith's reviewarXiv
  24. [44] Z. Shao, P. Wang, Q. Zhuet al., “Deepseekmath: Pushing the limits of mathematical reasoning in open languagemodels,”arXivpreprintarXiv:2402.03300,2024 on PithPith's reviewarXiv
  25. [45] A.Pichéetal.,“PipelineRL:Fasteron-policyreinforcementlearningforlongsequencegeneration,”arXiv preprintarXiv:2509.19128,2025 on PitharXiv
  26. [46] KimiTeam,“Kimik1.5: ScalingreinforcementlearningwithLLMs,”arXivpreprintarXiv:2501.12599,2025 on PithPith's reviewarXiv
  27. [47] A.Holtzman,J.Buys,L.Duetal.,“Thecuriouscaseofneuraltextdegeneration,”inInternationalConference onLearningRepresentations,2020. 140 BeamModelTechnicalReport,ReflectionTeametal on Pith
  28. [48] DeepSeek-AI,A.Liu,A.Meietal.,“Deepseek-v3.2: Pushingthefrontierofopenlargelanguagemodels,” 2025.[Online].Available: https://arxiv.org/abs/2512.02556 on PithPith's reviewarXiv
  29. [49] W.Maetal.,“StabilizingMoEreinforcementlearningbyaligningtrainingandinferencerouters,”arXiv preprintarXiv:2510.11370,2025 on PithPith's reviewarXiv
  30. [50] R.Greenblatt,A.Cotra,andH.Wijk,“Briefindependentinvestigationofagents’behavior,reasoningand collaborationintheOpenAI/HuggingFacehackingincident,”METRBlog,Aug.2026.[Online].Available: https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/ on Pith
  31. [51] I. Badertdinov, M. Nekrashevich, A. Shevtsovet al., “SWE-rebench V2: Language-Agnostic SWE Task Collection at Scale,” inProceedings of the 43rd International Conference on Machine Learning, ser. ProceedingsofMachineLearningResearch,vol.306. PMLR,2026,pp.4820–4843.[Online].Avai... on Pith
  32. [52] R. Cao, M. Chen, J. Chenet al., “Qwen3-Coder-Next Technical Report,” 2026. [Online]. Available: https://arxiv.org/abs/2603.00729 on PithPith's reviewarXiv
  33. [53] M.Chen,L.Zhang,Y.Fengetal.,“SWE-Universe: ScaleReal-WorldVerifiableEnvironmentstoMillions,” 2026.[Online].Available: https://arxiv.org/abs/2602.02361 on PithPith's reviewarXiv
  34. [54] J.Zhang,L.Ma,F.Wanetal.,“DockSmith: ScalingReliableCodingEnvironmentsviaanAgenticDocker Builder,”2026.[Online].Available: https://arxiv.org/abs/2602.00592 on PithPith's reviewarXiv
  35. [55] G.A.Oliva,G.K.Rajbahadur,A.Bhatiaetal.,“SPICE:AnAutomatedSWE-BenchLabelingPipelineforIssue Clarity,TestCoverage,andEffortEstimation,”2025.[Online].Available: https://arxiv.org/abs/2507.09108v5 on PitharXiv
  36. [56] J. Pan, X. Wang, G. Neubiget al., “Training Software Engineering Agents and Verifiers with SWE-Gym,” inProceedings of the 42nd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 267. PMLR, 2025, pp. 47717–47737. [Online]. Availabl... on Pith
  37. [57] N.Jain,J.Singh,M.Shettyetal.,“R2E-Gym: ProceduralEnvironmentGenerationandHybridVerifiersfor ScalingOpen-WeightsSWEAgents,”inSecondConferenceonLanguageModeling,2025.[Online].Available: https://openreview.net/forum?id=7evvwwdo3z on Pith
  38. [58] D.Zan, Z. Huang, W. Liuetal., “Multi-SWE-bench: AMultilingual Benchmarkfor Issue Resolving,” in AdvancesinNeuralInformationProcessingSystems,vol.38. CurranAssociates,Inc.,2025,datasetsand BenchmarksTrack.[Online].Available: https://proceedings.neurips.cc/paper_files/paper/2025... on Pith
  39. [59] C. Guo, J. Wu, S. Heet al., “MEnvAgent: Scalable Polyglot Environment Construction for Verifiable Software Engineering,” inProceedings of the 43rd International Conference on Machine Learning, ser. ProceedingsofMachineLearningResearch,vol.306. PMLR,2026,pp.38471–38495.[Online]... on Pith
  40. [60] J. Yang, K. Lieret, C. E. Jimenezet al., “SWE-smith: Scaling Data for Software Engineering Agents,” in AdvancesinNeuralInformationProcessingSystems,vol.38. CurranAssociates,Inc.,2025,datasetsand BenchmarksTrack.[Online].Available: https://proceedings.neurips.cc/paper_files/pap... on Pith
  41. [61] J.Zhao,G.Chen,F.Mengetal.,“ImmersionintheGitHubUniverse: ScalingCodingAgentstoMastery,” 2026.[Online].Available: https://arxiv.org/abs/2602.09892 on PitharXiv
  42. [62] M. A. Merrill, A. G. Shaw, N. Carliniet al., “Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces,” inInternational Conference on Learning Representations, 2026, pp. 40903–40986.[Online].Available: https://proceedings.iclr.cc/paper_files/pap... on Pith
  43. [63] Harbor Framework Team, “Harbor: A framework for evaluating and optimizing agents and models in containerenvironments,”Zenodo,2026.[Online].Available: https://doi.org/10.5281/zenodo.20953922 on PithPith's reviewdoi:10.5281/zenodo.20953922
  44. [64] Q.Yu,Z.Zhang,R.Zhuetal.,“DAPO:AnOpen-SourceLLMReinforcementLearningSystematScale,” 2025.[Online].Available: https://arxiv.org/abs/2503.14476 on PithPith's reviewarXiv
  45. [65] M. Luo, S. Tan, J. Wonget al., “DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Scaling RL,” NotionBlog,Feb.2025.[Online].Available: https://pretty-radio-b75.notion.site/DeepScaleR-Surpassing-O 1-Preview-with-a-1-5B-Model-by-Scaling-RL-19681902c1468005bed8ca303013a4e2 on Pith
  46. [66] I. Moshkov, D. Hanley, I. Sorokinet al., “AIMO-2 Winning Solution: Building State-of-the- Art Mathematical Reasoning Models with OpenMathReasoning dataset,” 2025. [Online]. Available: https://arxiv.org/abs/2504.16891 141 BeamModelTechnicalReport,ReflectionTeametal on PithPith's reviewarXiv
  47. [67] S.Alshammari,K.Wen,A.Zainaletal.,“MathNet: AGlobalMultimodalBenchmarkforMathematical ReasoningandRetrieval,”inInternationalConferenceonLearningRepresentations,2026,pp.12477–12507. [Online].Available: https://proceedings.iclr.cc/paper_files/paper/2026/hash/1516a7f7507d5550db5c7... on Pith
  48. [68] J. Dekoninck, N. Jovanović, T. Gehrungeret al., “Beyond Benchmarks: MathArena as an Evaluation PlatformforMathematicswithLLMs,”2026.[Online].Available: https://arxiv.org/abs/2605.00674 on PithPith's reviewarXiv
  49. [69] A.Meurer,C.P.Smith,M.Paprockietal.,“SymPy: symboliccomputinginPython,”PeerJComputerScience, vol.3,p.e103,Jan.2017.[Online].Available: https://doi.org/10.7717/peerj-cs.103 on Pithdoi:10.7717/peerj-cs.103
  50. [70] H.Kydlíček,“Math-Verify: MathVerificationLibrary,”Softwarerepository,2025,version0.6.1.[Online]. Available: https://github.com/huggingface/math-verify on Pith
  51. [71] Z. Cheng, S. Hao, T. Liuet al., “Revisiting Reinforcement Learning for LLM Reasoning from A Cross-DomainPerspective,”2025.[Online].Available: https://arxiv.org/abs/2506.14965 on PithPith's reviewarXiv
  52. [72] J. Cai, J. Zhu, R. Sunet al., “CodeContests-O: Powering LLMs via Feedback-Driven Iterative Test Case Generation,”2026.[Online].Available: https://arxiv.org/abs/2601.13682 on PitharXiv
  53. [73] Q. Shi, M. Tang, K. Narasimhanet al., “Can Language Models Solve Olympiad Programming?” 2024. [Online].Available: https://arxiv.org/abs/2404.10952 on PithPith's reviewarXiv
  54. [74] Z. Xia, K. Luo, H. Qianet al., “Open Data Synthesis For Deep Research,” 2025. [Online]. Available: https://arxiv.org/abs/2509.00375 on PithPith's reviewarXiv
  55. [75] Y.Du,R.Ye,S.Tangetal.,“OpenSeeker: DemocratizingFrontierSearchAgentsbyFullyOpen-Sourcing TrainingData,”2026.[Online].Available: https://arxiv.org/abs/2603.15594 on PitharXiv
  56. [76] R. Lu, Z. Hou, Z. Wanget al., “DeepDive: Advancing Deep Search Agents with Knowledge Graphs and Multi-TurnRL,”2025.[Online].Available: https://arxiv.org/abs/2509.10446 on PitharXiv
  57. [77] A.Gallant,“ripgrep,”Softwarerepository,2016.[Online].Available: https://github.com/BurntSushi/ripgrep on Pith
  58. [78] M.Senghaas,“GeneralAgent: ASelf-Evolving,SyntheticAgentEnvironment,”PrimeIntellectBlog,May 2026.[Online].Available: https://www.primeintellect.ai/blog/general-agent on Pith
  59. [79] Z.Xu,A.MezaSoria,S.Tanetal.,“TOUCAN:Synthesizing1.5MTool-AgenticDatafromReal-WorldMCP Environments,”2025.[Online].Available: https://arxiv.org/abs/2510.01179 on PitharXiv
  60. [80] NVIDIA, “Nemotron-SFT-Agentic-v2,” Hugging Face dataset, 2025. [Online]. Available: https: //huggingface.co/datasets/nvidia/Nemotron-SFT-Agentic-v2 on Pith
  61. [81] I. Gabriel and G. Keeling, “A matter of principle? AI alignment as the fair treatment of claims,”Philosophical Studies, vol. 182, no. 7, pp. 1951–1973, 2025. [Online]. Available: https: //link.springer.com/article/10.1007/s11098-025-02300-4 on Pithdoi:10.1007/s11098-025-02300-4
  62. [82] W. Zhao, X. Ren, J. Hesselet al., “Wildchat: 1m chatgpt interaction logs in the wild,”Proceedings of the Twelfth International Conference on Learning Representations, 2024. [Online]. Available: https://openreview.net/forum?id=Bl8u7ZRlbM on Pith
  63. [83] H.SheahanandM.Carroll,“Canpublicchatdatapredictreal-worldAImisalignments?” OpenAIAlignment ResearchBlog,Jun.2026.[Online].Available: https://alignment.openai.com/validating-public-evals/ on Pith
  64. [84] J. Engels, A. Conmy, B. Chughtaiet al., “SFT Drives Gemini’s Safety Properties,” LessWrong, Jun. 2026. [Online].Available: https://www.lesswrong.com/posts/nLrrYweeFxgXACSmS/sft-drives-gemini-s-safet y-properties-1 on Pith
  65. [85] F. Cairns, “SFT Also Drives Safety Eval Results in Olmo 3,” LessWrong, Sep. 2026. [Online]. Available: https://www.lesswrong.com/posts/d9rEnYcfCk4pn2PK8/sft-also-drives-safety-eval-results-in-olmo-3-1 on Pith
  66. [86] A.Wong,J.Engels,andN.Nanda,“SteeringRLTraining: BenchmarkingInterventionsAgainstReward Hacking,”LessWrong,Dec.2025.[Online].Available: https://www.lesswrong.com/posts/R5MdWGKsuvdP wGFBG/steering-rl-training-benchmarking-interventions-against on Pith
  67. [87] S. Marks, J. Lindsey, and C. Olah, “The Persona Selection Model: Why AI Assistants might Behave like Humans,” Anthropic Alignment Science Blog, Feb. 2026. [Online]. Available: https: //alignment.anthropic.com/2026/psm/ on Pith
  68. [88] C. Lu, J. Gallagher, J. Michalaet al., “The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models,” inProceedings of the 43rd International Conference on Machine Learning, ser. ProceedingsofMachineLearningResearch,vol.306. PMLR,2026,pp.80955–81003.[O... on Pith
  69. [89] A.T.Kalai,O.Nachum,S.S.Vempalaetal.,“WhyLanguageModelsHallucinate,”2025.[Online].Available: https://arxiv.org/abs/2509.04664 on PithPith's reviewarXiv
  70. [90] M. A. Mohamadi, T. Wang, and Z. Li, “Honesty over Accuracy: Trustworthy Language Models through ReinforcedHesitation,”2025.[Online].Available: https://arxiv.org/abs/2511.11500 on PithPith's reviewarXiv
  71. [91] P.Kirichenko,M.Ibrahim,K.Chaudhurietal.,“AbstentionBench: ReasoningLLMsFailonUnanswerable Questions,”inAdvancesinNeuralInformationProcessingSystems,vol.38. CurranAssociates,Inc.,2025. [Online].Available: https://proceedings.neurips.cc/paper_files/paper/2025/hash/fb122bfc3f0127... on Pith
  72. [92] D. Jackson, W. Keating, G. Cameronet al., “AA-Omniscience: Evaluating Cross-Domain Knowledge ReliabilityinLargeLanguageModels,”2025.[Online].Available: https://arxiv.org/abs/2511.13029 on PitharXiv
  73. [93] E. Choi, “GoodWiki Dataset,” Hugging Face dataset, Sep. 2023. [Online]. Available: https: //huggingface.co/datasets/euirim/goodwiki on Pith
  74. [94] A.Jacovi,A.Wang,C.Albertietal.,“TheFACTSGroundingLeaderboard: BenchmarkingLLMs’Abilityto GroundResponsestoLong-FormInput,”2025.[Online].Available: https://arxiv.org/abs/2501.03200 on PithPith's reviewarXiv
  75. [95] G. Dong, K. Lu, C. Liet al., “Self-play with Execution Feedback: Improving Instruction-following CapabilitiesofLargeLanguageModels,”inInternationalConferenceonLearningRepresentations,2025,pp. 39286–39313.[Online].Available: https://proceedings.iclr.cc/paper_files/paper/2025/ha... on Pith
  76. [96] Y.He,W.Li,H.Zhangetal.,“AdvancedIF:Rubric-BasedBenchmarkingandReinforcementLearningfor AdvancingLLMInstructionFollowing,”inProceedingsofthe64thAnnualMeetingoftheAssociationfor ComputationalLinguistics(Volume1: LongPapers). AssociationforComputationalLinguistics,2026,pp. 18003–... on Pith
  77. [97] Y. Bai, A. Jones, K. Ndousseet al., “Training a Helpful and Harmless Assistant with Reinforcement LearningfromHumanFeedback,”2022.[Online].Available: https://arxiv.org/abs/2204.05862 on PithPith's reviewarXiv
  78. [98] Y.Bai,S.Kadavath,S.Kunduetal.,“ConstitutionalAI:HarmlessnessfromAIFeedback,”2022.[Online]. Available: https://arxiv.org/abs/2212.08073 on PithPith's reviewarXiv
  79. [99] L.Gao,J.Schulman,andJ.Hilton,“ScalingLawsforRewardModelOveroptimization,”inProceedingsof the40thInternationalConferenceonMachineLearning,ser.ProceedingsofMachineLearningResearch,vol on Pith
  80. [100] J. Skalse, N. H. R. Howe, D. Krasheninnikovet al., “Defining and Characterizing Reward Hacking,” in AdvancesinNeuralInformationProcessingSystems,vol.35. CurranAssociates,Inc.,2022,pp.9460–9471. [Online].Available: https://proceedings.neurips.cc/paper_files/paper/2022/hash/3d71... on Pith

The first 100 of 199 references are listed here. All of them are in this paper's citation record.

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment