Pith.

Machine Learning · arXiv:2610.35059 v1 · submitted 26 Oct 2026 · 39 references

Reinforcement Learning Done Right for Explicit Confidence Targets

Zefeng He, Yichen He, Yuan Lin, Yang Yang, Yu Cheng

The paper says

“Large language models (LLMs) can produce plausible but incorrect answers when they are uncertain, which can mislead users and undermine model reliability.”

more of the abstract

Training models to abstain with reinforcement learning is a natural approach, but most existing methods use a fixed error penalty and therefore cannot adapt their behavior to different confidence requirements at inference time. We study explicit confidence targets, where a user specifies a target \(p\) in the prompt and the model directly chooses whether to answer or abstain, with the goal of achieving an accuracy of at least \(p\) among answered questions. However, learning to respond to different confidence targets with a single model has proven difficult, and prior attempts have failed to distinguish different targets. We show that this failure comes from two problems: the reward signal that distinguishes different values of p is often lost during optimization, while the original error penalty grows without bound as p approaches one. We propose RLEC (\textbfReinforcement \textbfLearning with \textbfExplicit \textbfConfidence Targets), which preserves the learning signal needed to distinguish confidence targets, introduces a stable reward family that keeps the desired answer condition unchanged, and uses reward scheduling to balance effective learning and training stability. Together, these designs enable the first successful training of a single LLM for explicit confidence target control. Extensive experiments across diverse benchmarks and model scales show that RLEC largely preserves general capabilities and achieves competitive calibration performance. RLEC also supports fine-grained target control without an external decision system or threshold calibration after training. By making explicit confidence targets part of the trained policy, RLEC opens a new and more direct path toward behavioral calibration.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 27 Oct 2026 · what was read was not recorded · not checked by a person

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

Holds up. Clear diagnosis of why prior abstention training failed, plus clean fixes that actually work for target-conditioned control. This line is wrong

6 minor

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

Are you an author? What you can do.

Not checked: figures, code, what the 39 references say.

6 minor objections

The paper says, at Abstract & Section 1

Pith says: minor objection 1 of 6

In the abstract and on page 2, 'ReinforcementLearning withExplicitConfidence Targets' appears without spaces (e.g., 'RLEC(ReinforcementLearning...'). Please insert spaces for readability.

The paper says, at Equation (9) and surrounding text, page 4

Pith says: minor objection 2 of 6

The LaTeX rendering of the expected reward expression is unclear: the term 'p1-p' appears without the fraction bar. The intended expression is likely (q(x)-p)/(1-p). Please ensure proper formatting.

show 4 more

The paper says, at Figure 3, page 7

Pith says: minor objection 3 of 6

The caption lists subfigures (a)–(d) but the text does not refer to specific subfigures when describing ERT and RLEC behavior. Adding explicit references (e.g., 'ERT ((a) and (c))' and 'RLEC ((b) and (d))') would improve clarity.

The paper says, at Table 3, Appendix B.1

Pith says: minor objection 4 of 6

The benchmark column 'OmniSci' is used without an earlier expansion. Adding a footnote or referring to the full name AA-Omniscience (as in Table 1) would help readers.

The paper says, at Various places, e.g. page 16

Pith says: minor objection 5 of 6

Several citations have missing spaces before the bracket (e.g., 'Damani et al.[8]', 'Ge et al.[9]'). Please ensure consistent spacing.

The paper says, at Appendix A.1, page 15

Pith says: minor objection 6 of 6

The phrase 'Clip-Higher' is capitalized but it is not clear if this is a proper name. It may be intended as 'clip-higher' or a specific technique; please clarify or use consistent terminology.

What would settle it

Run the Qwen3-4B-Base RLEC training with three different random seeds and report the mean and standard deviation of abstention rates at p=0.5 on the HotpotQA validation set after 500 steps. If the standard deviation exceeds 5 percentage points, the robustness of the target differentiation across training runs would be weaker, and the central claim would require qualification.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 27 Oct 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 46 s · 23,005 tokens read, 4,419 written
  3. Editor wrote the verdict above 15 s · 23,757 tokens read, 1,330 written
  4. Referee minor revision: 6 minor 36 s · 23,272 tokens read, 3,010 written

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

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This paper does what it says on the tin: it diagnoses why prior attempts to train a single LLM for explicit confidence targets failed, and then provides simple, well-motivated fixes that actually work. The core insight—that GRPO's standard-deviation normalization removes the p-dependent reward scale in the common case of binary-reward groups—is mathematically clean (Eq. 14) and empirically confirmed by the training dynamics plot in Figure 2. The proposed fixes are correspondingly straightforward: remove the std normalization, scale rewards with a family Rp(b) that preserves the theoretical decision rule q(x) > p, and linearly schedule b from 0 to 1 to balance early signal strength with later stability. The empirical story is convincing. On eight benchmarks at two model scales, RLEC produces clear separation across p values, while baselines (ERT, TruthRL, RLVR) remain flat. The ablation studies confirm each design choice individually, and the paper compares fairly against RLCR, the closest alternative formulation. The limitations section is unusually honest: the idealized guarantee assumes knowledge of q(x); on hard benchmarks like SimpleQA-Verified target violations persist at the 4B scale. The authors also note that limited compute prevented multiple training runs. These are real limitations, but they are openly stated and do not undermine the central claim that explicit target-conditioned abstention is feasible. The b-schedule is heuristic, though the paper tests a sinusoidal alternative and obtains comparable results, so this is a minor soft spot rather than a load-bearing flaw. This is a solid, well-executed paper. It makes a meaningful practical advance—turning a reported impossibility into a working system—and it explains why the prior approach failed. I would bring it to my reading group and cite it in my own work on selective prediction or RL for LLMs. It deserves a serious peer review.

The strongest objection

The paper's central claim—that RLEC enables the first successful training of a single LLM for explicit confidence target control—is well-supported by theoretical analysis, thorough ablations, and consistent empirical results across eight benchmarks and two model scales. The weakest assumption identified by the reader (that the reward family and linear b-schedule preserve the optimal decision rule while providing enough learning signal) is empirically validated through extensive ablation studies (Section 4.4, Appendix B.1) that compare fixed b values, the linear schedule, and a sinusoidal schedule. The paper transparently acknowledges its limitations: limited computational resources prevented multiple training runs, and the theoretical guarantee holds only under idealized assumptions (Section D). No internal inconsistency or critical hole in the argument is apparent. The empirical trends are strong and the causal mechanism (removal of standard-deviation normalization to preserve p-dependent reward scale) is directly supported by an ablation. Therefore, no additional load-bearing concern beyond those already acknowledged by the authors and the reader is identified.

The paper makes a clear central claim and supports it with a solid chain of reasoning: (1) identifying why prior methods fail (normalization loss and unbounded penalty), (2) proposing targeted fixes (mean-only centering, scaled reward family, b-scheduling), and (3) demonstrating successful target-conditioned abstention across two model scales. The ablations isolate each component and confirm the causal mechanism. Limitations are honestly discussed. The reader's weakest assumption—that the reward family and b-schedule preserve the decision rule while providing sufficient learning signal—is empirically validated, and the reader assigned ACCEPT with high confidence after weighing this. I concur. No new critical concern arises; therefore, the verdict should remain UNCHANGED.

The referee's report

If the results hold, this work makes a significant practical contribution by enabling a single LLM to adjust its abstention behavior on the fly according to a user-specified confidence target. This addresses a clear limitation of existing abstention methods that use fixed penalties or require external thresholding. The theoretical analysis of why standard-deviation normalization destroys the p-dependent signal is clean and supported by rollout-group composition measurements. The proposed reward family and scheduling are grounded in the analysis and validated through extensive ablations. The paper demonstrates the feasibility of behavioral calibration through direct policy training, opening a new direction for reliable LLM deployment.

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

Large language models can produce plausible but incorrect answers when uncertain. Rather than forcing a fixed trade-off between accuracy and coverage, this paper trains a single model to directly obey a confidence target p given in the prompt: answer only when at least p confident, otherwise abstain. Prior attempts (ERT) failed because the reinforcement learning signal that distinguishes different p values was lost by standard advantage normalization, and the error penalty grew unbounded as p approached 1. RLEC preserves the p-dependent reward signal by removing group standard deviation from advantage estimation, introduces a stable reward family parameterized by a coefficient b that bounds the penalty without changing the optimal decision rule, and schedules b from 0 to 1 over training to balance learning signal strength with stability. The result is the first successful training of a single LLM that smoothly varies its abstention rate with p, achieving accuracy among answered questions that generally meets or exceeds the requested target, with no external decision system required.

What carries the argument. The reward family $R_p^{{(b)}}$ = (1−p)^b * R_p, where R_p gives +1 for correct, −p/(1−p) for incorrect, and 0 for abstention. Multiplying all rewards at a given p by (1−p)^b scales the error penalty from unbounded (b=0) to bounded (b≥1) without changing the condition for answering (q(x) > p). Together with mean-centered advantages (removing group standard deviation), this preserves the p-dependent reward scale in the training signal. A linear schedule of b from 0 to 1 over training steps balances early strong signal for differentiation with later stability.

Core claim. A single LLM can be trained to follow explicit confidence targets (user-specified p in the prompt) if two design flaws in prior reinforcement learning are fixed: (1) the group standard deviation normalization in GRPO removes the p-dependent reward scale, collapsing advantages across different p values; (2) the original error penalty −p/(1−p) grows without bound near p=1, causing gradient dominance. RLEC removes the standard deviation normalization while keeping mean centering, uses a reward family $R_p^{{(b)}}$ = (1−p)^b * R_p that keeps the theoretical answer condition q(x) > p unchanged for all b, and linearly schedules b from 0 to 1. This yields a policy whose abstention rate rises steadily with p, with attempted accuracy generally at or above p, fine-grained control at 0.01 intervals, and extrapolation beyond the training range—all without external calibration or thresholding.

The premise it rests on. The reward family $R_p^{{(b)}}$ with a linear b-schedule preserves the theoretical optimal decision rule q(x) > p while providing a strong enough learning signal across all p values for the model to follow, an assumption validated only empirically on two model scales and not guaranteed theoretically when model capacity, optimization, and distribution shifts are considered.

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

A single LLM can be deployed to serve multiple reliability requirements simply by changing p in the prompt, eliminating the need for separate models or post-hoc threshold calibration.

RLEC achieves fine-grained control (p steps of 0.01) on difficult reasoning benchmarks like AIME, with smooth changes in abstention rate and accuracy across targets.

The method extrapolates to unseen p values beyond the training range (up to 0.95 when trained only to 0.90), suggesting that the learned policy generalizes the decision rule.

At larger scale (30B parameters), RLEC matches or exceeds the calibration performance of confidence-reporting methods (RLCR) while also providing direct answer/abstain decisions without external rules.

RLEC's reward family effectively turns a continuous confidence target problem into a family of related RL problems, where the stability coefficient b acts as a temperature on the reward that is p-conditional. This design principle—scaling rewards per target without altering the optimum—could apply to other RL domains with user-specified risk levels, such as autonomous driving or medical decision s

Because the method depends on the presence of ternary rollout groups (correct, incorrect, abstain) to transmit the p-signal, its effectiveness may degrade on tasks where the model rarely abstains or rarely answers correctly. The paper's ablations on simple vs. hard benchmarks hint at this boundary; a formal characterization of required group diversity would strengthen deployment guarantees.

The linear b-schedule is heuristic and may not be optimal. Future work could investigate adaptive schedules based on gradient variance or abstention rate dynamics, potentially improving stability on difficult knowledge-intensive benchmarks where target violations remain (e.g., SimpleQA-Verified).

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

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

The paper

Abstract

Large language models (LLMs) can produce plausible but incorrect answers when they are uncertain, which can mislead users and undermine model reliability. Training models to abstain with reinforcement learning is a natural approach, but most existing methods use a fixed error penalty and therefore cannot adapt their behavior to different confidence requirements at inference time. We study explicit confidence targets, where a user specifies a target \(p\) in the prompt and the model directly chooses whether to answer or abstain, with the goal of achieving an accuracy of at least \(p\) among answered questions. However, learning to respond to different confidence targets with a single model has proven difficult, and prior attempts have failed to distinguish different targets. We show that this failure comes from two problems: the reward signal that distinguishes different values of p is often lost during optimization, while the original error penalty grows without bound as p approaches one. We propose RLEC (\textbfReinforcement \textbfLearning with \textbfExplicit \textbfConfidence Targets), which preserves the learning signal needed to distinguish confidence targets, introduces a stable reward family that keeps the desired answer condition unchanged, and uses reward scheduling to balance effective learning and training stability. Together, these designs enable the first successful training of a single LLM for explicit confidence target control. Extensive experiments across diverse benchmarks and model scales show that RLEC largely preserves general capabilities and achieves competitive calibration performance. RLEC also supports fine-grained target control without an external decision system or threshold calibration after training. By making explicit confidence targets part of the trained policy, RLEC opens a new and more direct path toward behavioral calibration.

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 · 2026-10-27

The authors' main claim, as the classifier picked it out 0.91

“Together, these designs enable the first successful training of a single LLM for explicit confidence target control.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.92 · states limitation 0.25 · human participants 0.04 · controlled comparison 0.19 · randomized assignment 0.02 · states quantitative result 0.08 · compares with prior results 0.89 · mentions code or data availability 0.02

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
Large language models (LLMs) can produce plausible but incorrect answers when they are uncertain, which can mislead users and undermine model reliability. background 1.00 0.05 0.02 asserted 0.67
Training models to abstain with reinforcement learning is a natural approach, but most existing methods use a fixed error penalty and therefore cannot adapt their behavior to different confidence requirements at inference time. gap 0.72 0.05 0.03 asserted 0.99
We study \emph{explicit confidence targets}, where a user specifies a target \(p\) in the prompt and the model directly chooses whether to answer or abstain, with the goal of achieving an accuracy of at least \(p\) among answered questions. contribution 0.73 0.20 0.07 asserted 0.94
However, learning to respond to different confidence targets with a single model has proven difficult, and prior attempts have failed to distinguish different targets. gap 0.76 0.05 0.03 asserted 0.99
We show that this failure comes from two problems: the reward signal that distinguishes different values of $p$ is often lost during optimization, while the original error penalty grows without bound as $p$ approaches one. result 0.41 0.77 0.07 asserted 1.00
We propose RLEC (\textbf{R}einforcement \textbf{L}earning with \textbf{E}xplicit \textbf{C}onfidence Targets), which preserves the learning signal needed to distinguish confidence targets, introduces a stable reward family that keeps the desired answer condition unchanged, and uses reward scheduling to balance effective learning and training stability. method 0.76 0.87 0.14 asserted 1.00
Together, these designs enable the first successful training of a single LLM for explicit confidence target control. result 0.80 0.90 0.72 asserted 1.00
Extensive experiments across diverse benchmarks and model scales show that RLEC largely preserves general capabilities and achieves competitive calibration performance. result 1.00 0.95 0.16 asserted 1.00
RLEC also supports fine-grained target control without an external decision system or threshold calibration after training. result 0.93 0.90 0.10 asserted 1.00
By making explicit confidence targets part of the trained policy, RLEC opens a new and more direct path toward behavioral calibration. interpretation 0.99 0.50 0.25 asserted 0.68

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98

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, 2026-10-27: 1 request, 7,730 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 (10)
Figure 10 of the paper
Figure 10. Training dynamics of the abstention rate on HotpotQA under different choices of the stability coefficient and scheduling strategy. RLEC uses the linear schedule of b, while RLEC Sin. uses a sinusoidal schedule. Curves correspond to target confidence levels p ∈ {0.1, 0.5, 0.9}. Forced Accuracy. Among the configurations that distinguish confidence targets, Forced Accuracy remains broadly stable, with the linear schedule obtaining the highest aggregate Forced Accuracy. The clearest differences appear on AIME, where fixed b = 0, fixed b = 0.5, and LR Schedule are all more than three…

Showing 1 of the paper's 10 figures. Figures are the authors', reproduced from the paper.

References (39)
  1. [1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023 on PithPith's reviewarXiv
  2. [2] Michael Bereket and Jure Leskovec. Uncalibrated reasoning: Grpo induces overconfidence for stochastic outcomes. arXiv preprint arXiv:2508.11800, 2025 on PithPith's reviewarXiv
  3. [3] Aili Chen, Aonian Li, Bangwei Gong, Binyang Jiang, Bo Fei, Bo Yang, Boji Shan, Changqing Yu, Chao Wang, Cheng Zhu, et al. Minimax-m1: Scaling test-time compute efficiently with lightning attention.arXiv preprint arXiv:2506.13585, 2025 on PithPith's reviewarXiv
  4. [4] Hui Chen, Antoine Didisheim, and Luciano A Somoza. Out of the black box: Uncertainty quantification for llms via conditional probabilities. Technical report, National Bureau of Economic Research, 2026 on Pith
  5. [5] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021 on PithPith's reviewarXiv
  6. [6] Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabil... on PithPith's reviewarXiv
  7. [7] Yuyang Dai and Yuxia Wang. Rescaling confidence: What scale design reveals about llm metacognition.arXiv preprint arXiv:2603.09309, 2026 on PithPith's reviewarXiv
  8. [8] Mehul Damani, Isha Puri, Stewart Slocum, Idan Shenfeld, Leshem Choshen, Yoon Kim, and Jacob Andreas. Beyond binary rewards: Training lms to reason about their uncertainty. InInternational conference on learning representations, volume 2026, pages 59456–59493, 2026 on Pith
  9. [9] Cheng Ge, Caitlyn Heqi Yin, Hao Liang, and Jiawei Zhang. Why grpo needs normalization: A local-curvature perspective on adaptive gradients.arXiv preprint arXiv:2601.23135, 2026 on PithPith's reviewarXiv
  10. [10] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025 on PithPith's reviewarXiv
  11. [11] Lukas Haas, Gal Yona, Giovanni D’Antonio, Sasha Goldshtein, and Dipanjan Das. Simpleqa verified: A reliable factuality benchmark to measure parametric knowledge.arXiv preprint arXiv:2509.07968, 2025 on PitharXiv
  12. [12] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset.arXiv preprint arXiv:2103.03874, 2021 on PithPith's reviewarXiv
  13. [13] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.ACM transactions on information... on Pith
  14. [14] Muhammed Emrullah Ildiz, Halil Alperen Gozeten, Ege Onur Taga, and Samet Oymak. Learning to correct: Calibrated reinforcement learning for multi-attempt chain-of-thought.arXiv preprint arXiv:2604.17912, 2026 on PithPith's reviewarXiv
  15. [15] Declan Jackson, William Keating, George Cameron, and Micah Hill-Smith. Aa-omniscience: Evaluating cross- domain knowledge reliability in large language models.arXiv preprint arXiv:2511.13029, 2025 on PitharXiv
  16. [16] Abha Jha, Akanksha Mahajan, Ashwath Vaithinathan Aravindan, Praveen Saravanan, Sai Sailaja Policharla, and Sonal Chaturbhuj Gehlot. Rewarding intellectual humility learning when not to answer in large language models. arXiv preprint arXiv:2601.20126, 2026 on PitharXiv
  17. [17] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation.ACM computing surveys, 55(12):1–38, 2023 on Pith
  18. [18] Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221, 2022 on PithPith's reviewarXiv
  19. [19] Adam Tauman Kalai, Ofir Nachum, Santosh S Vempala, and Edwin Zhang. Why language models hallucinate. arXiv preprint arXiv:2509.04664, 2025 on PithPith's reviewarXiv
  20. [20] Amita Kamath, Robin Jia, and Percy Liang. Selective question answering under domain shift. InProceedings of the 58th annual meeting of the association for computational linguistics, pages 5684–5696, 2020. 13 on Pith
show 19 more
  1. [21] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. InInternational Conference on Learning Representations, volume 2024, pages 39578–39601, 2024 on Pith
  2. [22] Stephanie Lin, Jacob Hilton, and Owain Evans. Teaching models to express their uncertainty in words.arXiv preprint arXiv:2205.14334, 2022 on PithPith's reviewarXiv
  3. [23] Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. Understanding r1-zero-like training: A critical perspective.arXiv preprint arXiv:2503.20783, 2025 on PithPith's reviewarXiv
  4. [24] Mohamad Amin Mohamadi, Tianhao Wang, and Zhiyuan Li. Honesty over accuracy: Trustworthy language models through reinforced hesitation.arXiv preprint arXiv:2511.11500, 2025 on PithPith's reviewarXiv
  5. [25] Muyu Pan, Shu Zhao, Nan Zhang, Philip Shin, Varun Parekh, Vijaykrishnan Narayanan, and Rui Zhang. Tiar: Trajectory-informed advantage reweighting for llm abstention learning.arXiv preprint arXiv:2605.25850, 2026 on PithPith's reviewarXiv
  6. [26] David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark.arXiv preprint arXiv:2311.12022, 2023 on PithPith's reviewarXiv
  7. [27] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024 on PithPith's reviewarXiv
  8. [28] Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework. InProceedings of the Twentieth European Conference on Computer Systems, pages 1279–1297, 2025 on Pith
  9. [29] Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, Jianfeng Cai, Xinyuan Cai, Peizhou Cao, Yuxuan Cao, Ziwei Chai, Y Charles, et al. Kimi k3: Open frontier intelligence.arXiv preprint arXiv:2607.24653, 2026 on PitharXiv
  10. [30] Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D Manning. Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback. InProceedings of... on Pith
  11. [31] Benjamin Turtel, Danny Franklin, Kris Skotheim, Luke Hewitt, and Philipp Schoenegger. Outcome-based reinforcement learning to predict the future.arXiv preprint arXiv:2505.17989, 2025 on PitharXiv
  12. [32] Qingni Wang, Yue Fan, and Xin Wang. Safer: Risk-constrained sample-then-filter in large language models. In International Conference on Learning Representations, volume 2026, pages 53723–53747, 2026 on Pith
  13. [33] Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark.Advances in Neural Information Processing Systems, 37:95... on Pith
  14. [34] Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models.arXiv preprint arXiv:2411.04368, 2024 on PithPith's reviewarXiv
  15. [35] Zhepei Wei, Xiao Yang, Kai Sun, Jiaqi Wang, Rulin Shao, Jingxiang Chen, Mohammad Kachuee, Teja Gollapudi, Yiwei Liao, Nicolas Scheffer, et al. Truthrl: Incentivizing truthful llms via reinforcement learning.arXiv preprint arXiv:2509.25760, 2025 on PithPith's reviewarXiv
  16. [36] Jiayun Wu, Jiashuo Liu, Zhiyuan Zeng, Tianyang Zhan, Tianle Cai, and Wenhao Huang. Mitigating llm hallucination via behaviorally calibrated reinforcement learning.arXiv preprint arXiv:2512.19920, 2025 on PitharXiv
  17. [37] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025 on PithPith's reviewarXiv
  18. [38] Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. InProceedings of the 2018 conference on empirical methods in natural language process... on Pith
  19. [39] Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. Dapo: An open-source llm reinforcement learning system at scale.Advances in Neural Information Processing Systems, 38:113222–113244, 2026. 14 Appen... on Pith
Cited by 1 Pith paper

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