Pith. sign in

REVIEW 4 major objections 4 minor 35 references

RTTC: Reward-Guided Collaborative Test-Time Compute

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A pretrained reward model can route each query to direct inference, retrieval-augmented generation, or test-time training, and this per-query routing beats applying any single strategy uniformly.

desk verdict Reward-guided routing between RAG and TTT is a genuinely new idea, but the paper's own tables contradict the 'consistently superior' claim and the reward signal is never validated. read the letter →

arxiv 2508.10024 v1 pith:343TWSIV submitted 2025-08-07 cs.CL cs.AIcs.IR

J. Pablo Muñoz , Jinjie Yuan This is my paper
classification cs.CLcs.AIcs.IR
keywords test-timecomputereward-guidedroutingretrieval-augmentedgenerationtrainingLoRAfine-tuningquery-statecachingadaptiveinferenceLLMadaptation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Test-time compute methods like retrieval-augmented generation (RAG) and test-time training (TTT) improve LLM answers but cost compute, and neither wins for every query. RTTC claims that a pretrained reward model can serve as a per-query router: score the direct answer; if it clears a threshold, stop; otherwise retrieve samples and try RAG, and only if RAG's answer outscores the direct one, fine-tune the model on those samples and answer. Across four instruction-tuned LLMs and five coding, math, and medical benchmarks, the routed system outperforms applying RAG or TTT uniformly, and running both and picking the higher-reward answer does better still. The paper also introduces query-state caching, which reuses retrieved samples and fine-tuned adapters across similar queries, cutting redundant computation with near-unchanged accuracy. The stake is that inference-time compute budgets can be spent where they help.

What carries the argument

The load-bearing object is the reward-guided decision gate: a pretrained reward model R that assigns a scalar score R(x, y) to a query-answer pair, plus two comparisons — r0 ≥ τr for early exit and r_RAG > r0 for preferring RAG — followed by a fallback to TTT on the same retrieved samples. The second mechanism is query-state caching: an embedding-similarity threshold τe decides whether to reuse cached retrieved samples (for RAG) or cached low-rank adapters (for TTT) from a similar historical query, with a budget and eviction policy to bound the cache.

What would settle it

For a sample of queries, run all three strategies and record both reward scores and actual correctness. If a substantial share of queries has r_RAG > r_0 while the RAG answer is no more accurate than the direct answer, or r_0 ≥ τr while the direct answer is wrong and TTT would have fixed it, the reward signal is not carrying the routing decision.

Watch

Extended reading notes

Core claim

The central claim is that strategy selection at test time can be driven by a reward model evaluating the model's own outputs. The pipeline is: generate a direct response and score it; if the score is at or above a fixed threshold, return it; otherwise retrieve similar examples from a knowledge base, generate a RAG response, and keep it only if its reward exceeds the direct response's; otherwise fine-tune the base model on those examples via low-rank adapters and return that response. In experiments this reward-guided cascade improves average accuracy over both vanilla RAG and vanilla TTT on every tested model, with the largest gains on the joint variant that runs RAG and TTT in parallel and

Load-bearing premise

The routing depends on a single pretrained reward model's scalar scores tracking true answer quality across coding, math, and medical questions, and on a hand-set threshold; if either fails, queries get sent to the wrong branch.

Editorial extensions

If this is right

  • Compute budgets become query-adaptive: easy queries cost one forward pass plus one reward score, while hard queries pay for retrieval and fine-tuning only when earlier stages fail.
  • The joint variant, which runs RAG and TTT in parallel and takes the higher-reward answer, consistently beats the sequential cascade, suggesting that more candidate strategies could be added to the same gate.
  • Query-state caching can amortize test-time compute across similar queries: the reported cache hit rates exceed 60% for both RAG samples and TTT adapters, with accuracy within 97-100% of the uncached system.
  • The same reward-guided gate extends to any test-time compute strategy, not only RAG and TTT, so the framework is a general way to choose among inference-time adapters.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same router may transfer to unseen domains without retraining, but only if a 0.6B reward model's calibration holds; a cheap check is to measure routing agreement with an oracle that knows which strategy actually helps on a held-out domain.
  • Editorial inference: the threshold τr acts as a compute-accuracy dial; because the appendix shows gains flatten or costs rise as τr increases, a learned or per-query threshold could tighten the trade-off that the paper leaves manual.
  • Editorial inference: shadow deployment is the natural next test — run RTTC alongside a static strategy on live traffic and log reward scores, routed branches, and downstream outcomes; the routing's real-world value depends on reward scores tracking performance outside benchmark subsets.
  • Editorial inference: the current prototype's server can recover user prompts, so the architecture assumes a privacy bridge; routing on embeddings or encrypted query states would be a direct extension preserving the reward decisions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes RTTC, a reward-guided framework that selects per query among direct inference, retrieval-augmented generation (RAG), and test-time training (TTT), using a pretrained reward model both as an absolute gate (threshold tau_r) and as a relative comparator (r_RAG > r_0). The system is deployed as a server-client architecture with a multi-domain knowledge base, and the paper introduces Query-State Caching (QSC) to reuse retrieved samples and trained LoRA adapters across similar queries. Experiments are reported on four instruction-tuned LLMs (Llama-3-8B, Llama-3.1-8B, Mistral-7B, Qwen2.5-3B) across five tasks (MBPP, HumanEval, MathQA, GSM-Plus, ATC), with additional ablations on retrieval count, reward threshold, and reward model size. The central claim is that RTTC 'consistently' outperforms vanilla RAG and TTT, and that reward-guided selection is necessary.

Significance. If the central claim were established, the paper would make a useful practical contribution: it unifies three test-time strategies under one decision rule, provides a distributed system design, and includes a cost model (Table 4) that makes the overhead of each branch explicit. The ablations in Appendices C and D are also in the right direction: they examine how routing changes with threshold and reward model, and the paper is candid about limitations such as manual hyperparameters and the optimistic cache-hit measurement. However, the empirical evidence as presented does not support the headline claim. The paper's own Table 3 contains a direct counterexample (Mistral-7B), the main comparison selects the best retrieval sample size on the test set with no error bars, and the reward model's calibration—the load-bearing assumption of the routing mechanism—is never validated. As a result, the current evidence is not strong enough to justify the claimed necessity of reward-guided selection.

major comments (4)
  1. [§4.2, Table 3] The abstract and Section 4.2 claim that RTTC 'consistently achieves superior accuracy' over RAG and TTT. This is contradicted by the Mistral-7B results in Table 3: RTTC averages 29.3 versus TTT's 29.9, and RTTC-Joint averages 29.9, which is at best a tie. Per-task, RTTC is below TTT on MBPP (32.4 vs 38.4) and HumanEval (37.2 vs 38.4). Since one of the four base models directly contradicts the headline, the 'consistent' claim is not supported by the reported data.
  2. [§4.1, Table 3 footnote, Appendix B] The main comparison is based on 'the best performance achieved across the evaluated retrieval sample sizes {1,2,4,8,16} for each method,' selected on the test set, and no variance or significance information is reported. Several tasks use small subsets (200 samples for MathQA and GSM-Plus, 600 for ATC). This selection on the test set makes the reported improvements optimistic and does not support a claim of general superiority. Appendix Tables 8–11 show many retrieval sizes where RTTC is not the best: for example, in Table 8 (Llama-3-8B) at k=8, RTTC averages 39.1 versus TTT's 40.8, and at k=16, RTTC averages 39.4 versus TTT's 41.4. The paper should report a fixed protocol, error bars, and a statistical comparison.
  3. [§3.1, Appendix D, Table 13] The routing mechanism assumes that the pretrained Skywork-Reward-V2-0.6B model's scores are usable quality signals across coding, math, and medical tasks, both as an absolute gate (tau_r=2.0) and as a relative comparison (r_RAG > r_0). No evidence is presented that reward scores correlate with task correctness. Appendix D, Table 13 shows that swapping the reward model changes the no-adaptation share from 1.2% to 38.5% (Llama-3-8B, tau_r=5) while average accuracy stays essentially flat (41.7 vs 42.0). This is compatible with the routing decisions being only weakly tied to correctness. The paper should report a reward-correctness calibration analysis (e.g., accuracy stratified by reward score, or an oracle-routing upper bound) before claiming that reward-guided selection is the cause of the observed gains.
  4. [§3.3, Algorithm 2] Algorithm 2 uses an inner-product similarity metric gamma (Section 4.1) but computes ex* = argmin over gamma(ex_i, ex_t) and then reuses the cache when gamma(ex*, ex_t) > tau_e. For an inner-product similarity, the most similar cached state corresponds to the argmax, not the argmin. As written, a cache hit requires the least similar historical embedding to exceed the threshold, which is inconsistent with the reported 62–70% RAG/TTT cache hit rates in Table 6. Please correct the pseudocode or clarify the actual similarity/selection convention used in the implementation.
minor comments (4)
  1. [Algorithm 1, Step 3] In Algorithm 1, line 9 computes rRAG = R(x, y_RAG), but the augmented input x' is used for generation; the reward evaluation should use the same augmented input, i.e., R(x', y_RAG), unless the notation is intentionally simplified.
  2. [§4.4, Table 6] The paper acknowledges that cache utilization may be inflated by evaluating benchmark samples from the same domain sequentially. This is a useful caveat, but the table still reports only point estimates; an error bar or a simulation with shuffled query order would help quantify the effect.
  3. [Figure 1] The y-axis label 'Accuracy Improvement (%)' is unclear because values exceed 100 and are relative to the no-adaptation baseline. Consider labeling as 'Relative accuracy (%)' or plotting the actual average accuracy.
  4. [§4.6 / Appendix D] The appendix tables are dense and would benefit from a column indicating whether the reported number is the best over retrieval sizes or a fixed retrieval size; Table 6 explicitly uses k=4 while Table 3 uses best-over-k, which can confuse readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reward model and retrieval/TTT procedures are external inputs, and the decision rule is not constructed from its own outputs.

full rationale

RTTC's central mechanism is a pretrained reward model (Skywork-Reward-V2) that scores an initial response and, when below a manually set threshold, routes to RAG or TTT. Nothing in Algorithm 1 or Section 3.1 defines the predicted accuracy in terms of the reward scores; the reward model is a fixed external artifact, not fitted to the downstream benchmarks or to RTTC's own outputs. The threshold tau_r=2.0 and the retrieval sample-size grid {1,2,4,8,16} are experimental choices, and the Limitations section explicitly acknowledges the manual threshold as a limitation; this is a tuning/robustness concern, not a definitional circularity. Appendix D's reward-model swap shows routing changes with reward-model choice while average accuracy stays roughly flat, which bears on whether reward scores track correctness, but it does not make the derivation circular. The paper reports the best result over retrieval sample sizes and includes a case (Mistral-7B, RTTC 29.3 vs TTT 29.9) that weakens the 'consistently superior' claim; these are empirical/correctness issues outside the circularity definition. No load-bearing self-citation or uniqueness theorem is invoked. The framework is therefore self-contained with respect to circularity: any weaknesses are about reward-model calibration and evaluation protocol, not about the claimed result being built into its inputs.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The framework's effectiveness rests on several borrowed components and hand-set parameters. The reward model and embedding model are pretrained and not validated on these tasks. The threshold tau_r, cache threshold tau_e, budget b, and TTT hyperparameters are manually chosen. The main table selects the best retrieval sample size per method, which is a post-hoc choice. No new physical entities are introduced.

free parameters (6)
  • reward threshold tau_r = 2.0 (default), varied 2.0/5.0/8.0
    Hand-set; controls when adaptation triggers. Ablation in Appendix C shows strong effect on accuracy and strategy distribution.
  • retrieval sample count k = best of {1,2,4,8,16} per method
    Main results report best k per method, a post-hoc selection on test data (note under Table 3).
  • QSC reuse threshold tau_e = 0.5
    Section 4.1; reuse similarity threshold.
  • QSC budget b = 8
    Section 4.1; cache size.
  • LoRA rank / alpha = 32 / 16
    Section 4.1; TTT adapter configuration.
  • TTT learning rate = 5e-5
    Section 4.1; two epochs, batch size 1.
assumptions (6)
  • domain assumption Pretrained reward models provide reliable signals of response quality across coding, math, and medical Q&A.
    Used in Section 3.1 Steps 1 and 3 and Section 4.1; the entire routing decision depends on r0 and rRAG being comparable, but no calibration or per-task analysis is provided.
  • domain assumption Retrieved samples from the knowledge base are relevant enough to improve RAG and TTT.
    Section 3.1 Steps 2 and 4 and Section 4.5; retrieval quality is only qualitatively shown via heatmap and t-SNE.
  • domain assumption LoRA fine-tuning on retrieved samples generalizes to the query's task.
    Section 3.1 Step 4; TTT success depends on this, using a single hyperparameter set.
  • domain assumption The evaluation subsets (200 or 600 samples) are representative of the full benchmarks.
    Section 4.1 and Table 2; no confidence intervals provided.
  • domain assumption The knowledge base contains sufficient in-domain coverage for the evaluation tasks.
    Appendix A; constructed from datasets related to (but not identical to) the benchmarks.
  • domain assumption Similarity metric gamma with inner product can identify reusable query states.
    Section 3.3 and Algorithm 2; QSC effectiveness depends on embedding similarity correspondence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RTTC: Reward-Guided Collaborative Test-Time Compute." pith.science (2026). https://pith.science/paper/343TWSIV

@misc{pith2026250810024,
  author       = {Pith},
  title        = {Pith review of: RTTC: Reward-Guided Collaborative Test-Time Compute},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/343TWSIV}},
  note         = {Machine review of arXiv:2508.10024}
}
read the original abstract

Test-Time Compute (TTC) has emerged as a powerful paradigm for enhancing the performance of Large Language Models (LLMs) at inference, leveraging strategies such as Test-Time Training (TTT) and Retrieval-Augmented Generation (RAG). However, the optimal adaptation strategy varies across queries, and indiscriminate application of TTC strategy incurs substantial computational overhead. In this work, we introduce Reward-Guided Test-Time Compute (RTTC), a novel framework that adaptively selects the most effective TTC strategy for each query via a pretrained reward model, maximizing downstream accuracy across diverse domains and tasks. RTTC operates in a distributed server-client architecture, retrieving relevant samples from a remote knowledge base and applying RAG or lightweight fine-tuning on client devices only when necessary. To further mitigate redundant computation, we propose Query-State Caching, which enables the efficient reuse of historical query states at both retrieval and adaptation levels. Extensive experiments across multiple LLMs and benchmarks demonstrate that RTTC consistently achieves superior accuracy compared to vanilla RAG or TTT, validating the necessity of adaptive, reward-guided TTC selection and the potential of RTTC for scalable, high-performance language model adaptation.

Figures

Figures reproduced from arXiv: 2508.10024 by the authors.

Figure 1
Figure 1. Performance overview of RTTC on down￾stream tasks across various LLMs. “Accuracy Improve￾ment” indicates the relative average score on five down￾stream tasks evaluated in §4. TTC Strategy Latency Memory Accuracy No Adaptation (Direct Inference) + + + Retrieval-Augmented Generation (RAG) ++ ++ +++ Test-Time Training (TTT) +++ +++ +++ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Workflow of Reward-guided Test-Time Compute (RTTC). For each query, a pretrained reward model evaluates candidate responses and selects the optimal adaptation strategy (No Adaptation, RAG, or TTT). τr is a predefined threshold. additional fine-tuning steps and memory for model updates. Moreover, naively applying RAG or TTT to every query can lead to unnecessary computa￾tion and inefficient resource utilization. A su… view at source ↗
Figure 3
Figure 3. Overview of the retrieval and test-time compute stages. When the reward model determines that the LLM [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Overview of the Query-State Caching (QSC) strategy for multi-turn caching and efficient test-time compute. QSC is compatible with both RAG and TTT: for RAG, the cached state consists of retrieved samples; for TTT, the cached state stores the fine-tuned model state (e.g…
Figure 5
Figure 5. Figure 5: Domain distribution of retrieval samples from [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: t-SNE visualization of the embeddings for [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 7 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Ekin Aky \"u rek, Mehul Damani, Linlu Qiu, Han Guo, Yoon Kim, and Jacob Andreas. 2024. The surprising effectiveness of test-time training for abstract reasoning. arXiv preprint arXiv:2411.07279

    arXiv 2024

  4. [4]

    Aida Amini, Saadia Gabriel, Shanchuan Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. 2019. https://doi.org/10.18653/v1/N19-1245 M ath QA : Towards interpretable math word problem solving with operation-based formalisms . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics:...

    doi:10.18653/v1/n19-1245 2019

  5. [5]

    ARC-AGI. 2025. A R C P rize --- arcprize.org. https://arcprize.org. [Accessed 13-03-2025]

    work page 2025

  6. [6]

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and 1 others. 2021. Program synthesis with large language models. arXiv preprint arXiv:2108.07732

    arXiv 2021

  7. [7]

    Loubna Ben Allal, Niklas Muennighoff, Logesh Kumar Umapathi, Ben Lipkin, and Leandro von Werra. 2022. A framework for the evaluation of code generation models. https://github.com/bigcode-project/bigcode-evaluation-harness

    work page 2022

  8. [8]

    Bradley Brown, Jordan Juravsky, Ryan Saul Ehrlich, Ronald Clark, Quoc V Le, Christopher Re, and Azalia Mirhoseini. 2025. https://openreview.net/forum?id=0xUEBQV54B Large language monkeys: Scaling inference compute with repeated sampling

    work page 2025

Show all 35 references
  1. [9]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey W...

    2020

  2. [10]

    Junying Chen, Zhenyang Cai, Ke Ji, Xidong Wang, Wanlong Liu, Rongsheng Wang, Jianye Hou, and Benyou Wang. 2024. https://arxiv.org/abs/2412.18925 Huatuogpt-o1, towards medical complex reasoning with llms . Preprint, arXiv:2412.18925

    2024 arXiv

  3. [11]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott G...

    2021 arXiv

  4. [12]

    Fran c ois Chollet. 2019. https://arxiv.org/abs/1911.01547 On the measure of intelligence . CoRR, abs/1911.01547

    2019 arXiv

  5. [13]

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. https://arxiv.org/abs/2401.08281 The faiss library

    2024 arXiv

  6. [14]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  7. [15]

    Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang S...

    2023 doi

  8. [16]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. 2023 b . Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2(1)

    2023 arXiv

  9. [17]

    Moritz Hardt and Yu Sun. 2024. Test-time training on nearest neighbors for large language models. In International Conference on Learning Representations

    2024

  10. [18]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

    2022

  11. [19]

    Jonas H \"u botter, Sascha Bongni, Ido Hakimi, and Andreas Krause. 2024. Efficiently learning at test-time: Active fine-tuning of llms. arXiv preprint arXiv:2410.08020

    2024 arXiv

  12. [20]

    IronbarArc24. 2024. arc24 --- ironbar.github.io. https://ironbar.github.io/arc24/. [Accessed 13-03-2025]

    2024

  13. [21]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, and 1 others. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825

    2023 arXiv

  14. [22]

    Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. 2020. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. arXiv preprint arXiv:2009.13081

    2020 arXiv

  15. [23]

    Denis Kocetkov, Raymond Li, Loubna Ben Allal, Jia Li, Chenghao Mou, Carlos Muñoz Ferrandis, Yacine Jernite, Margaret Mitchell, Sean Hughes, Thomas Wolf, Dzmitry Bahdanau, Leandro von Werra, and Harm de Vries. 2022. The stack: 3 tb of permissively licensed source code. Preprint

    2022

  16. [24]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural informati...

    2020

  17. [25]

    Qintong Li, Leyang Cui, Xueliang Zhao, Lingpeng Kong, and Wei Bi. 2024. https://arxiv.org/abs/2402.19255 Gsm-plus: A comprehensive benchmark for evaluating the robustness of llms as mathematical problem solvers . Preprint, arXiv:2402.19255

    2024 arXiv

  18. [26]

    Chris Yuhao Liu, Liang Zeng, Yuzhen Xiao, Jujie He, Jiacai Liu, Chaojie Wang, Rui Yan, Wei Shen, Fuxiang Zhang, Jiacheng Xu, Yang Liu, and Yahui Zhou. 2025. Skywork-reward-v2: Scaling preference data curation via human-ai synergy. arXiv preprint arXiv:2507.01352

    2025 arXiv

  19. [27]

    Arindam Mitra, Hamed Khanpour, Corby Rosset, and Ahmed Awadallah. 2024. https://arxiv.org/abs/2402.14830 Orca-math: Unlocking the potential of slms in grade school math . Preprint, arXiv:2402.14830

    2024 arXiv

  20. [28]

    Ankit Pal, Logesh Kumar Umapathi, and Malaikannan Sankarasubbu. 2022. https://proceedings.mlr.press/v174/pal22a.html Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering . In Proceedings of the Conference on Health, Inference, and Lea...

    2022

  21. [29]

    Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2025. https://openreview.net/forum?id=4FWAwZtd2n Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning . In The Thirteenth International Conference on Learning Representations

    2025

  22. [30]

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. 2020. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229--9248. PMLR

    2020

  23. [31]

    Chi, Quoc V Le, and Denny Zhou

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. 2022. https://openreview.net/forum?id=_VjQlMeSB_J Chain of thought prompting elicits reasoning in large language models . In Advances in Neural Information Proc...

    2022

  24. [32]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, and 1 others. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671

    2024 arXiv

  25. [33]

    Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. 2023. Metamath: Bootstrap your own mathematical questions for large language models. arXiv preprint arXiv:2309.12284

    2023 arXiv

  26. [34]

    Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. 2025. Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176

    2025 arXiv

  27. [35]

    Hu Zhiqiang, Lan Yihuai, Wang Lei, Xu Wanyu, Lim EePeng, Lee Roy Ka-Wei, Bing Lidong, and Poria Soujanya. 2023. Llm-adapters: An adapter family for parameter-efficient fine-tuning of large language models. arXiv preprint arXiv:2304.01933

    2023 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.