Pith.

Artificial Intelligence · arXiv:2610.23085 v1 · submitted 16 Oct 2026 · 42 references

Rethinking GRPO for Large Reasoning Model Safety Alignment

Jiawei Chen, Yi Zhang, Jiawei Du, Zhonghou Lv, Tairan Chen, Junyuan Shang, Jinchang Hou, Xiao Yang

show 4 more authors

Yu Tian, Chenfu Bao, Chao Yu, Zhaoxia Yin

The paper says

“Large reasoning models (LRMs) exhibit strong reasoning capabilities but remain vulnerable to unsafe outputs. While Group Relative Policy Optimization (GRPO) is widely used for safety alignment, its group-relative optimization may provide insufficient safety-directed learning signals.”

more of the abstract

In this paper, we identify two failure modes: advantage collapse, where homogeneous rollout groups yield vanishing advantages, and unsafe diversification, where unsafe behaviors persist and become increasingly diverse. Our theoretical analysis reveals that behavioral consistency weakens within-group safety contrasts, while entropy regularization can diversify unsafe responses when relative signals vanish. To address these challenges, we propose Beacon, comprising Safety Signal Recovery (SSR) and Safety-Utility Decoupled Alignment (SUDA). SSR combines advantage-guided trajectory filtering and safe reference supervision to recover effective safety learning signals, while SUDA employs category-aware rewards and benign-only KL anchoring to mitigate over-refusal without compromising safety. Experiments across four Qwen3 and DeepSeek-R1-Distill variants and eight benchmarks show that our framework achieves the best average safety and reasoning performance, reducing average safety error by up to 32.83 percentage points while maintaining low over-refusal.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 19 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. Solid theoretical diagnosis of GRPO's safety alignment failures, packaged with a well-engineered fix that works across models. This line is wrong

2 major objections · 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 42 references say.

The paper says, at Tables 1 and 2

The review did not quote this passage.

Pith says: Objection 1 of 2 · major

The paper reports single-point performance without confidence intervals or statistical significance tests. Given that training involves randomness from sampling and optimization, variance estimates (e.g., across multiple seeds or bootstrap intervals on test sets) would strengthen the reproducibility claims and help readers assess the reliability of the reported improvements. I recommend adding error bars or a statement about the number of training runs conducted.

The paper says, at Figure 6

The review did not quote this passage.

Pith says: Objection 2 of 2 · major

The sensitivity analysis for κ is informative, but the y-axis labels show 'Change relative to κ=1.00' without specifying the direction (e.g., safety error reduction in percentage points). The caption says 'change relative to κ=1.00' but it would be clearer to explicitly state that negative values indicate improvement for safety and positive for reasoning.

6 minor objections

The paper says, at Abstract, line 12

Pith says: minor objection 1 of 6

'average safety error by up to 32.83 percentage points while maintaining low over-refusal' — consider adding a brief mention that over-refusal is measured on XSTest for completeness.

The paper says, at Theorem 1, Eq. (5)

Pith says: minor objection 2 of 6

The derivation assumes i.i.d. rollouts from the current policy at a single optimization step. While this is standard for per-step analysis, explicitly noting that the analysis is per-step (and may change over training) would avoid confusion. The text after Eq. (6) already discusses implications; a brief qualifier in the theorem statement would help.

show 4 more

The paper says, at Table 2, caption

Pith says: minor objection 3 of 6

The table header 'Safety (↓)' is ambiguous: does it include both harmful response rates and over-refusal? The column lists 'Avg.' which averages the four safety metrics; clarifying the averaging method (e.g., simple arithmetic mean of the four rates) in the caption would help.

The paper says, at Section 4.2.1, Eq. (13)

Pith says: minor objection 4 of 6

The notation 'Top-κ (P_q; Δ_i)' is used where κ is a fraction; consider defining κ as the retention ratio (0 < κ ≤ 1) to avoid confusion with the number of pairs (κG/2). The text already says 'top-κ fraction', but adding a brief note in the equation caption would improve clarity.

The paper says, at Appendix D, Algorithm 1

Pith says: minor objection 5 of 6

In line 4, the loop variable i is used for rollouts; later in line 15, pairs are formed. It would be clearer to use separate indices (e.g., i for rollouts, k for pairs) to avoid ambiguity.

The paper says, at Figures 3, 4, 5

Pith says: minor objection 6 of 6

The font sizes in the legend and axes are very small; please ensure that the figures are readable when printed at standard journal column width.

What would settle it

Independently replicate the Qwen3-8B experiment using an alternative safe reference dataset (e.g., self-generated via a different safety-aligned model) and a different reward model (e.g., a DeBERTa-based reward). If the safety improvements shrink by more than 5 absolute percentage points relative to the reported 1.88% average safety error, then performance is more dependent on the specific data and reward model than the framework's architecture would suggest.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 19 Oct 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 1 min 6 s · 26,029 tokens read, 4,837 written
  3. Editor wrote the verdict above 14 s · 26,877 tokens read, 1,165 written
  4. Referee minor revision: 2 major objections, 6 minor 40 s · 26,352 tokens read, 2,857 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 good safety alignment work should: it identifies a real failure mode in a widely used method, explains why it happens, and builds a targeted fix. The core insight—that GRPO's group-relative signal vanishes when responses become behaviorally consistent—is simple once stated, but nobody has spelled it out this clearly for safety. Theorem 1 is almost trivial in retrospect but it matters, because it forces practitioners to think about what happens when their rollout groups are all-safe or all-unsafe. Theorem 2 on unsafe diversification under entropy regularization is also well-taken.

The resulting Beacon framework is sensible and technically clean. The four components (AGTF, SRS, CAR, BKL) each address a specific piece of the problem, and the ablations show each one pulls its weight. The experimental evaluation is unusually thorough: five model scales, two families, eight benchmarks, multiple evaluators including human annotation. The safety gains (17–33 percentage points reduction in safety error) are large and consistent, and the over-refusal numbers stay low. That's a genuine achievement.

The main soft spot is the dependence on the ArmoRM reward model and ThinkSafe reference data. The authors acknowledge this in Section 8, so it's not a hidden flaw, but it does mean the framework's performance may shift with a different reward model or safety data source. The consistency check across Llama-Guard-3, GPT-6, and human evaluators helps, but those are safety classifiers, not reward models. A follow-up experiment swapping the reward model would strengthen the claim of generalizability.

Code and data are not released, which limits reproducibility but is fairly standard at this stage.

Overall: the paper's theoretical grounding is sound, the empirical support is strong, and the limitations are honestly stated. This is a serious contribution to LRM safety alignment.

My recommendation: accept for peer review. The paper deserves a thoughtful referee process and will likely hold up well.

The strongest objection

The paper presents a well-motivated theoretical analysis of GRPO failure modes in LRM safety alignment and proposes a practical framework (Beacon) with four complementary components. The theoretical claims (Theorems 1 and 2) are internally consistent and correctly characterize the limitations of group-relative optimization under behavioral consistency. The empirical evaluation is extensive: five model scales (0.6B–14B), two model families, eight benchmarks, and multiple evaluators including human annotation. Ablations confirm each component's contribution. The authors explicitly acknowledge the main limitation—dependence on reward model quality and safe reference data—and provide evidence that results are consistent across evaluators (Llama-Guard-3, GPT-6, human). No internal contradiction, unsupported leap, or hidden assumption was found that would undermine the central claim that Beacon improves safety while preserving reasoning and limiting over-refusal. The paper's strengths (cross-model consistency, robust ablations, theoretical grounding) outweigh the acknowledged dependence on external signals.

The reader identified the dependence on reward model and safe reference data as the weakest assumption. The authors acknowledge this limitation and provide consistency checks across evaluators that partially mitigate it. My stress-test found no additional load-bearing concern that would alter the verdict. The theoretical analysis is sound under its stated assumptions, the experiments are thorough and well-controlled, and the ablations support the contribution of each component. While the paper does not release code or data, this is not a correctness issue. The central claim—that Beacon improves safety alignment while maintaining reasoning and limiting over-refusal—is well-supported by the evidence presented. Therefore, no adjustment to the ACCEPT verdict is warranted.

The referee's report

If the results hold, the paper makes several important contributions: (i) it provides a formal characterization of why GRPO can fail to provide useful safety gradients under behavioral consistency, (ii) it proposes a practical framework that demonstrably mitigates these failure modes, and (iii) it offers extensive empirical evidence across multiple model families, scales, and benchmarks. The theoretical analysis is rigorous under its stated assumptions, and the experiments include robust ablations, adversarial attack evaluation, and cross-evaluator consistency checks. The framework's reliance on external reward models is clearly acknowledged. The work is likely to influence how GRPO-based safety alignment is designed for LRMs.

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

The paper argues that Group Relative Policy Optimization (GRPO), a popular method for aligning large AI models, suffers from two fundamental problems when used to make reasoning models safer. First, as models become more consistent, the safety differences within sampled groups vanish, causing GRPO's learning signal to drop to zero—a phenomenon called advantage collapse. Second, when no contrast exists, the entropy regularizer that normally encourages exploration instead spreads probability across diverse unsafe responses, making the problem worse. The authors propose Beacon, a framework that recovers safety signals by filtering for high-contrast rollouts and adding safe reference supervision, while decoupling safety and utility objectives to avoid over-refusal. Across multiple models and benchmarks, Beacon reduces safety errors by up to 33 percentage points while maintaining or improving reasoning accuracy.

What carries the argument. The key theoretical machinery is Theorem 1, which shows that the probability of a group containing both safe and unsafe responses ($P_{\text{mix}} = 1 - p_q^G - (1-p_q)^G$) vanishes as the policy becomes safe-consistent ($p_q \to 1$) or unsafe-consistent ($p_q \to 0$). In the unsafe-consistent regime, all responses receive identical rewards, so GRPO's advantages are zero and the objective reduces to regularizers alone. Theorem 2 then shows that under these conditions entropy regularization flattens the unsafe distribution ($H(\pi^*) \ge H(\pi_{\text{ref}})$), explaining unsafe diversification. The practical machinery is Beacon, built on Safety Signal Recovery (SSR)—advantage-guided trajectory filtering plus safe reference supervision—and Safety-Utility Decoupled Alignment (SUDA)—category-aware rewards and benign-only KL anchoring.

Core claim. The central discovery is that GRPO's group-relative advantage computation is structurally prone to failure in safety alignment of large reasoning models. As post-training increases behavioral consistency, the probability that a sampled group contains both safe and unsafe responses—the condition for a non-zero relative safety signal—approaches zero. When this happens, GRPO provides no gradient direction toward safer behavior, and the entropy regularizer, lacking any safety-directed signal, flattens the distribution over unsafe responses. The paper shows that these two failure modes, advantage collapse and unsafe diversification, are not accidental but follow from the mathematics of GRPO combined with the nature of safety-relevant responses in reasoning models. Beacon addresses them by directly recovering safety signals through trajectory filtering and safe reference supervision, and by treating harmful and benign prompts differently to prevent over-refusal.

The premise it rests on. Beacon's effectiveness depends on the quality and reliability of the reward model (ArmoRM-Llama3-8B) and the safe reference responses (from ThinkSafe); if either is biased or flawed, the claimed gains may not generalize.

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

If Beacon's analysis is correct, any safety alignment method that relies solely on group-relative rewards (like vanilla GRPO) will systematically fail on prompts where the model is consistently unsafe, because those prompts produce zero gradient.

Beacon's safe reference supervision (SRS) component can be seen as a drop-in fix for GRPO-trained models, requiring only paired safe references for harmful prompts.

The decoupled alignment strategy (separate reward formulations for harmful vs. benign prompts) provides a concrete way to reduce the safety tax—degraded performance on legitimate requests—without sacrificing safety.

Beacon's success across model sizes (0.6B to 14B) suggests the failure modes are scale-invariant, implying that larger models will also benefit from signal recovery.

Theorem 1 implies that GRPO's utility for safety alignment is inherently bounded by the model's initial behavior: if a model starts consistently unsafe on a prompt, GRPO alone cannot learn safety from its own rollouts. An editor might hypothesize that pre-training a small classification head to detect such prompts and then applying targeted supervision would be a lighter alternative to full SRS.

The unsafe diversification result suggests that entropy bonuses, commonly used to prevent mode collapse in RL, actively harm safety alignment when safety signals vanish. One could test whether replacing entropy regularization with a different exploration bonus (e.g., count-based or curiosity-driven) avoids spreading mass over unsafe responses.

Beacon relies on the quality of safe references and reward models; an extension could use a judge model to generate safe references on the fly, removing the need for a curated dataset. This would broaden applicability to domains without pre-written safe responses.

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 reasoning models (LRMs) exhibit strong reasoning capabilities but remain vulnerable to unsafe outputs. While Group Relative Policy Optimization (GRPO) is widely used for safety alignment, its group-relative optimization may provide insufficient safety-directed learning signals. In this paper, we identify two failure modes: advantage collapse, where homogeneous rollout groups yield vanishing advantages, and unsafe diversification, where unsafe behaviors persist and become increasingly diverse. Our theoretical analysis reveals that behavioral consistency weakens within-group safety contrasts, while entropy regularization can diversify unsafe responses when relative signals vanish. To address these challenges, we propose Beacon, comprising Safety Signal Recovery (SSR) and Safety-Utility Decoupled Alignment (SUDA). SSR combines advantage-guided trajectory filtering and safe reference supervision to recover effective safety learning signals, while SUDA employs category-aware rewards and benign-only KL anchoring to mitigate over-refusal without compromising safety. Experiments across four Qwen3 and DeepSeek-R1-Distill variants and eight benchmarks show that our framework achieves the best average safety and reasoning performance, reducing average safety error by up to 32.83 percentage points while maintaining low over-refusal.

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 · 2026-10-20

The authors' main claim, as the classifier picked it out 1.00

“Experiments across four Qwen3 and DeepSeek-R1-Distill variants and eight benchmarks show that our framework achieves the best average safety and reasoning performance, reducing average safety error by up to 32.83 percentage points while maintaining low over-refusal.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.11 · human participants 0.03 · controlled comparison 0.58 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.83 · mentions code or data availability 0.03

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 reasoning models (LRMs) exhibit strong reasoning capabilities but remain vulnerable to unsafe outputs. background 1.00 0.08 0.02 asserted 1.00
While Group Relative Policy Optimization (GRPO) is widely used for safety alignment, its group-relative optimization may provide insufficient safety-directed learning signals. gap 0.52 0.10 0.05 hedged possibility 0.99
In this paper, we identify two failure modes: advantage collapse, where homogeneous rollout groups yield vanishing advantages, and unsafe diversification, where unsafe behaviors persist and become increasingly diverse. contribution 0.52 0.88 0.11 asserted 1.00
Our theoretical analysis reveals that behavioral consistency weakens within-group safety contrasts, while entropy regularization can diversify unsafe responses when relative signals vanish. interpretation 0.47 0.91 0.08 asserted 1.00
To address these challenges, we propose Beacon, comprising Safety Signal Recovery (SSR) and Safety-Utility Decoupled Alignment (SUDA). contribution 0.74 0.85 0.11 asserted 1.00
SSR combines advantage-guided trajectory filtering and safe reference supervision to recover effective safety learning signals, while SUDA employs category-aware rewards and benign-only KL anchoring to mitigate over-refusal without compromising safety. method 0.95 0.83 0.10 asserted 1.00
Experiments across four Qwen3 and DeepSeek-R1-Distill variants and eight benchmarks show that our framework achieves the best average safety and reasoning performance, reducing average safety error by up to 32.83 percentage points while maintaining low over-refusal. result 1.00 0.96 0.80 asserted 1.00

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

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, 2026-10-20: 1 request, 5,681 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 (5)
Figure 1 of the paper
Figure 1. Two failure modes of vanilla GRPO we re
Figure 2 of the paper
Figure 2. Overview of our proposed framework for effective LRM safety alignment. Safety Signal Recovery (SSR)
Figure 3 of the paper
Figure 3. Comparison of GRPO variants throughout training. (a) Advantage collapse: proportion of all￾unsafe groups with near-zero advantages. (b) Unsafe persistence: proportion of all-unsafe groups. (c) Unsafe diversification: semantic entropy of unsafe responses. KL anchoring, which constrains the policy toward its pre-alignment behavior only on benign prompts. Specifically, for q ∈ Db, we further augment the GRPO objective with an explicit KL constraint: Jb(θ) = JGRPO(θ) − λbDKL (πθ(·|q)∥πbase(·|q)), (17) where πbase denotes the policy before safety post￾training and λb controls the…
Figure 5 of the paper
Figure 5. Robustness comparison against four adversar

Showing 4 of the paper's 5 figures. Figures are the authors', reproduced from the paper.

References (42)
  1. [1] Reducing the Safety Tax in LLM Safety Alignment with On-Policy Self-Distillation , author=. 2026 , eprint= on Pith
  2. [2] Star-1: Safer alignment of reasoning llms with 1k data , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume= on Pith
  3. [3] Thinksafe: Self-generated safety alignment for reasoning models , author=. arXiv preprint arXiv:2601.23143 , year= on PithPith's reviewarXiv
  4. [4] Rafailov, Rafael and Sharma, Archit and Mitchell, Eric and Manning, Christopher D and Ermon, Stefano and Finn, Chelsea , booktitle =. Direct Preference Optimization: Your Language Model is Secretly a Reward Model , url =. doi:10.52202/075280-2338 , editor = on Pithdoi:10.52202/075280-2338
  5. [5] Zhang, Yichi and Ding, Yue and Yang, Jingwen and Luo, Tianwei and Li, Dongbai and Duan, Ranjie and Liu, Qiang and Su, Hang and Dong, Yinpeng and Zhu, Jun , booktitle =. Towards Safe Reasoning in Large Reasoning Models via Corrective Intervention , url = on Pith
  6. [6] NGRPO: Negative-enhanced Group Relative Policy Optimization , author=. 2025 , eprint= on Pith
  7. [7] Advantage Collapse in Group Relative Policy Optimization: Diagnosis and Mitigation , author=. Forty-third International Conference on Machine Learning , year= on Pith
  8. [8] Le, Thanh-Long V. and Jeon, Myeongho and Vu, Kim and Lai, Viet and Yang, Eunho , booktitle =. No Prompt Left Behind: Exploiting Zero-Variance Prompts in LLM Reinforcement Learning via Entropy-Guided Advantage Shaping , url = on Pith
  9. [9] GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization , author=. 2026 , eprint= on Pith
  10. [10] EDGE-GRPO: Entropy-Driven GRPO with Guided Error Correction for Advantage Diversity , author=. 2025 , eprint= on Pith
  11. [11] Zhang, Xichen and Wu, Sitong and Zhu, Yinghao and Tan, Haoru and Yu, Shaozuo and He, Ziyi and Jia, Jiaya , booktitle =. Scaf-GRPO: Scaffolded Group Relative Policy Optimization for Enhancing LLM Reasoning , url = on Pith
  12. [12] Zhou, Kaiwen and Zhao, Xuandong and Srinivasa, Jayanth and Liu, Gaowen and Feng, Aosong and Song, Dawn and Wang, Xin Eric. S afe K ey: Amplifying Aha-Moment Insights for Safety Reasoning. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025 on Pith
  13. [13] Dineen, Jacob and Rrv, Aswin and Liu, Qin and Xu, Zhikun and Ye, Xiao and Shen, Ming and Li, Zhaonan and Lu, Shijie and Baral, Chitta and Chen, Muhao and Zhou, Ben. QA - LIGN : Aligning LLM s through Constitutionally Decomposed QA. Findings of the Association for Computational... on Pith
  14. [14] Competitive programming with large reasoning models , author=. arXiv preprint arXiv:2502.06807 , year= on PithPith's reviewarXiv
  15. [15] Guo, Daya and Yang, Dejian and Zhang, Haowei and Song, Junxiao and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Zhang, Ruoyu and Ma, Shirong and Bi, Xiao and Zhang, Xiaokang and Yu, Xingkai and Wu, Yu and Wu, Z. F. and Gou, Zhibin and Shao, Zhihong and Li, Zhuoshu and Gao, Zi... on Pith
  16. [16] Proximal Policy Optimization Algorithms , author=. 2017 , eprint= on Pith
  17. [17] DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint= on Pith
  18. [18] Qwen3 Technical Report , author=. 2025 , eprint= on Pith
  19. [19] Jiang, Liwei and Rao, Kavel and Han, Seungju and Ettinger, Allyson and Brahman, Faeze and Kumar, Sachin and Mireshghallah, Niloofar and Lu, Ximing and Sap, Maarten and Choi, Yejin and Dziri, Nouha , booktitle =. WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarial... on Pithdoi:10.52202/079017-1493
  20. [20] Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonable , author=. 2025 , eprint= on Pith
show 22 more
  1. [21] Jiang, Fengqing and Xu, Zhangchen and Li, Yuetai and Niu, Luyao and Xiang, Zhen and Li, Bo and Lin, Bill Yuchen and Poovendran, Radha. S afe C hain: Safety of Language Models with Long Chain-of-Thought Reasoning Capabilities. Findings of the Association for Computational Lingu... on Pithdoi:10.18653/v1/2025.findings-acl.1197
  2. [22] Jeung, Wonje and Sangyeon, Yoon and Kahng, Minsuk and No, Albert , booktitle =. SAFEPATH: Preventing Harmful Reasoning in Chain-of-Thought via Early Alignment , url =. doi:10.52202/085713-3331 , editor = on Pithdoi:10.52202/085713-3331
  3. [23] Salmani-Zarchi, Mohammad Mahdi and Rahimi, Zahra and Faili, Heshaam and Dousti, Mohammad Javad. MDP - GRPO : Stabilized Group Relative Policy Optimization for Multi-Constraint Instruction Following. Proceedings of the 64th Annual Meeting of the A ssociation for C omputational ... on Pithdoi:10.18653/v1/2026.acl-long.1982
  4. [24] Training Verifiers to Solve Math Word Problems , author=. 2021 , eprint= on Pith
  5. [25] Lightman, Hunter and Kosaraju, Vineet and Burda, Yuri and Edwards, Harrison and Baker, Bowen and Lee, Teddy and Leike, Jan and Schulman, John and Sutskever, Ilya and Cobbe, Karl , booktitle =. Let s Verify Step by Step , url = on Pith
  6. [26] American Invitational Mathematics Examination (AIME) 2024 , author= on Pith
  7. [27] GPQA: A Graduate-Level Google-Proof Q&A Benchmark , author=. 2023 , eprint= on Pith
  8. [28] Souly, Alexandra and Lu, Qingyuan and Bowen, Dillon and Trinh, Tu and Hsieh, Elvis and Pandey, Sana and Abbeel, Pieter and Svegliato, Justin and Emmons, Scott and Watkins, Olivia and Toyer, Sam , booktitle =. A StrongREJECT for Empty Jailbreaks , url =. doi:10.52202/079017-398... on Pithdoi:10.52202/079017-3984
  9. [29] HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal , author=. 2024 , eprint= on Pith
  10. [30] R. XST est: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024. doi:10... on Pithdoi:10.18653/v1/2024.naacl-long.301
  11. [31] The Llama 3 Herd of Models , author =. 2024 , eprint = on Pith
  12. [32] Han, Seungju and Rao, Kavel and Ettinger, Allyson and Jiang, Liwei and Lin, Bill Yuchen and Lambert, Nathan and Choi, Yejin and Dziri, Nouha , booktitle =. WildGuard: Open One-stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs , url =. doi:10.52202/079017... on Pithdoi:10.52202/079017-0261
  13. [33] Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts , author=. EMNLP , year= on Pith
  14. [34] Lambert, Nathan and Pyatkin, Valentina and Morrison, Jacob and Miranda, LJ and Lin, Bill Yuchen and Chandu, Khyathi and Dziri, Nouha and Kumar, Sachin and Zick, Tom and Choi, Yejin and Smith, Noah A. and Hajishirzi, Hannaneh. R eward B ench: Evaluating Reward Models for Langua... on Pithdoi:10.18653/v1/2025.findings-naacl.96
  15. [35] Efficient Safety Alignment of Language Models via Latent Personality Traits , author=. arXiv preprint arXiv:2607.07918 , year= on PithPith's reviewarXiv
  16. [36] NovaSky Team , title= on Pith
  17. [37] Reasoning as an adaptive defense for safety , author=. Advances in Neural Information Processing Systems , volume= on Pith
  18. [38] RAPO: Risk-Aware Preference Optimization for Generalizable Safe Reasoning , author=. arXiv preprint arXiv:2602.04224 , year= on PitharXiv
  19. [39] Chen, Jiawei and Yang, Yang and Yu, Chao and Tian, Yu and Cao, Zhi and Yang, Xue and Li, Linghao and Su, Hang and Yin, Zhaoxia. Red Teaming Large Reasoning Models. Proceedings of the 64th Annual Meeting of the A ssociation for C omputational L inguistics (Volume 1: Long Papers). 2026 on Pith
  20. [40] Chen, Jiawei and Yang, Xiao and Fang, Zhengwei and Tian, Yu and Dong, Yinpeng and Yin, Zhaoxia and Su, Hang. A uto B reach: Universal and Adaptive Jailbreaking with Efficient Wordplay-Guided Optimization via Multi- LLM s. Findings of the A ssociation for C omputational L ingui... on Pith
  21. [41] Liu, Shuyuan and Chen, Jiawei and Ruan, Shouwei and Su, Hang and Yin, Zhaoxia , title =. Proceedings of the 32nd ACM International Conference on Multimedia , pages =. 2024 , isbn =. doi:10.1145/3664647.3680616 , abstract = on Pithdoi:10.1145/3664647.3680616
  22. [42] Tex3D: Objects as Attack Surfaces via Adversarial 3D Textures for Vision-Language-Action Models , author=. 2026 , eprint= 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