arxiv: 2605.11467 · v1 · submitted 2026-05-12 · 💻 cs.LG · cs.AI

Recognition: no theorem link

Drop the Act: Probe-Filtered RL for Faithful Chain-of-Thought Reasoning

Swapnil Parekh

Authors on Pith no claims yet

Pith reviewed 2026-05-13 02:03 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords chain-of-thought reasoningreinforcement learningreasoning faithfulnesspost-commitment detectionGRPOprobe filteringreasoning theater

0 comments

The pith

A probe trained once on frozen model activations filters unfaithful post-commitment steps during reinforcement learning to reduce reasoning theater.

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

The paper shows that reasoning models often produce chains of thought that rationalize answers they have already committed to internally, creating unhelpful theater that wastes tokens and obscures true computation. ProFIL trains a multi-head attention probe once on the frozen base model using verifier-derived labels to detect these post-commitment steps from activations alone. During GRPO rollouts, high probe-score trajectories have their advantages zeroed, which suppresses theater without requiring human annotation or retraining the base model. This yields shorter, more faithful chains while preserving accuracy across multiple domains and model sizes, and it outperforms simple length penalties by targeting semantic commitment rather than length alone.

Core claim

A probe trained on a frozen base with verifier-derived labels and no human annotation provides a stable signal that suppresses theater while resisting the RL-obfuscation failure mode predicted by prior work. Across GSM8K, LiveCodeBench, ToolUse, and MMLU-Redux with Llama-8B and Qwen-7B, ProFIL reduces post-commitment theater by 11-100%, raises faithful fraction by up to 24 percentage points under an independent judge, shortens chains by 4-19%, and preserves or improves task accuracy.

What carries the argument

The multi-head attention probe that detects post-commitment steps from internal activations alone, with advantage zeroing applied to high-scoring GRPO rollouts.

If this is right

Post-commitment theater drops by 11 to 100 percent across four domains.
Faithful chain fraction increases, for example by 24 points on LiveCodeBench under an external judge.
Average chain length falls 4 to 19 percent while task accuracy stays the same or rises.
The gains exceed those from a matched length-penalty baseline, showing the effect is not just compression.
The probe signal works without human labels and resists the obfuscation mode seen in prior RL work.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same commitment-detection idea could be inserted into other reinforcement-learning loops for reasoning models beyond GRPO.
Internal activations appear to carry detectable commitment information early enough that a one-time probe suffices for ongoing training.
Models trained this way may converge toward naturally shorter and more faithful chains rather than learning to hide theater.
The released probes and rollouts allow direct testing of whether the filter generalizes to new model families or tasks.

Load-bearing premise

A probe trained once on frozen activations can reliably detect post-commitment steps across rollouts, and zeroing advantages for those trajectories suppresses theater without introducing new biases or degrading the policy.

What would settle it

If running ProFIL on a held-out reasoning task shows no reduction in post-commitment steps detected by an independent judge or produces a measurable drop in final task accuracy, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2605.11467 by Swapnil Parekh.

**Figure 1.** Figure 1: Reasoning theater before and after ProFIL (perf-ratio 0.60 → 0.00, 6 post-commit steps eliminated). A randomly-sampled high-performativity GSM8K rollout (performativity ratio > 0.9) from DeepSeek-R1-Distill-Llama-8B under standard GRPO (left) and the paired rollout under ProFIL (right) on the same question. Each row is one reasoning step; color encodes the frozen-base probe’s per-step performativity score … view at source ↗

**Figure 2.** Figure 2: ProFIL turns a frozen-base probe into a training signal. (1) Forced-answering yields verifier-derived step-level performativity labels (no human annotation): the first step whose forced answer is correct defines the commitment point; later steps are performative. (2) A gated multi-head attention probe (RLFR-style; 8) is trained once on activations of the frozen base at layers ℓ1, ℓ2, then never updated. (3… view at source ↗

**Figure 3.** Figure 3: ProFIL suppresses theater while preserving or improving accuracy on all four domains. Left: performativity ratio drops 11–100% in every domain (lower is better), with non-overlapping CIs on GSM8K and LiveCodeBench. Right: task accuracy is preserved or improved in every case. Probe AUROC exceeds 0.92 in all domains ( [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Length compression is not the operative mechanism (LiveCodeBench full test set, all three conditions). Length penalty increases performativity vs. baseline (0.267 → 0.374); ProFIL reduces it by 72% (→ 0.076) and raises faithful fraction from 0.733 to 0.924, with chain length shrinking by 507 chars vs. baseline. The key distinction is semantic: ProFIL eliminates postcommitment steps, whereas the length-pen… view at source ↗

**Figure 5.** Figure 5: Theater wastes tokens and introduces bugs; ProFIL eliminates both. Highest-contrast case among 14 high-theater baseline examples (LiveCodeBench: max odd–even frequency difference). Baseline (9,883 chars, perf = 0.995): circular reasoning revisits the same edge case three times before writing correct code, deliberative-looking but contributing nothing. Length Penalty (4,886 chars, perf = 0.962): shorter bu… view at source ↗

**Figure 6.** Figure 6: Frozen base is validated; inference-time steering is insufficient. Left: the student probe agrees with the frozen-base probe on baseline and ProFIL rollouts, validating the frozen-base measurement. Right: the best CAA steering coefficient matches the unsteered ProFIL model and remains above ProFIL’s evaluation performativity (dashed). Training is the operative mechanism. 0.0 0.2 0.4 0.6 0.8 1.0 Per-rollout… view at source ↗

**Figure 7.** Figure 7: Theater has a concrete linguistic signature that ProFIL eliminates (LiveCodeBench, Llama-8B). Left: distribution of per-rollout performativity ratios. The baseline is bimodal: rollouts are either near-faithful or highly theatrical. Length penalty shifts the mean only slightly; ProFIL collapses the high-theater mode entirely. Center: fraction of rollouts containing structural rederivation markers (### Appr… view at source ↗

**Figure 8.** Figure 8: Theater suppression persists across all chain lengths, ruling out length compression as the mechanism (LiveCodeBench). Rollouts are stratified into 5 length quintiles on the pooled distribution; the panel reports performativity ratio within each bin. ProFIL suppresses theater across the full length distribution, including the longest 20% of responses, while also producing shorter chains overall (11,875 vs.… view at source ↗

**Figure 9.** Figure 9: Adaptive τ eliminates manual threshold tuning. Left: the proposed linear schedule maps mean group accuracy to τ ∈ [0.20, 0.50]; dots mark the three LCB checkpoints. Right: filter rates under fixed τ=0.5, fixed τ=0.2, and adaptive τ at each checkpoint. At step 320 (lowest accuracy) adaptive τ automatically matches the sparse-regime value; at step 440 it rises, balancing signal preservation and theater filte… view at source ↗

**Figure 10.** Figure 10: Theater marks success, not uncertainty: it is a learned confidence ritual. Left: task accuracy by performativity tercile (T1 = low theater, T3 = high theater) on LiveCodeBench baseline rollouts. High-theater rollouts trend toward higher accuracy: the high-theater tercile averages ∼33% accuracy vs. ∼9% for the low-theater tercile (Spearman ρ=+0.240, p=0.016; AUROC = 0.33, meaning theater anti-predicts fail… view at source ↗

**Figure 11.** Figure 11: Step-by-step forced-answer trajectories for two self-correcting GSM8K rollouts. Each [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗

read the original abstract

Reasoning models post-hoc rationalize answers they have already committed to internally, producing chains of *reasoning theater*: deliberative-looking steps that contribute nothing to correctness. This wastes inference tokens, pollutes interpretability, and obscures what the model actually computed. We introduce **ProFIL** (**Pro**be-**Fil**tered Reinforcement Learning) to *reduce theater, increase chain-of-thought faithfulness, and shrink chain length* in a single, drop-in extension to Group Relative Policy Optimization (GRPO). A multi-head attention probe is trained *once* on the *frozen* base model to detect post-commitment steps from internal activations alone; during GRPO, rollouts whose probe score exceeds a threshold have their advantage zeroed. *Our central finding is that a probe trained on a frozen base, with verifier-derived labels and no human annotation, provides a stable signal that suppresses theater while resisting the RL-obfuscation failure mode predicted by prior work.* Across four reasoning domains (GSM8K, LiveCodeBench, ToolUse, MMLU-Redux) and two model architectures (Llama-8B, Qwen-7B), ProFIL reduces post-commitment theater by **11--100%**, raises faithful-fraction (e.g., +24pp on LiveCodeBench under an independent Claude 3.7 Sonnet judge), and shortens chains by 4--19%, all while preserving or improving task accuracy. ProFIL also beats a matched length-penalty GRPO baseline, isolating the gain as semantic commitment-detection rather than chain compression. Probe weights, training configurations, and rollouts are released across all four domains.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

ProFIL adds a frozen probe to GRPO that zeros advantages on high-theater rollouts, delivering measurable cuts in post-commitment steps across domains while beating length penalties.

read the letter

The core contribution is training a multi-head attention probe once on frozen base-model activations using verifier labels, then filtering GRPO trajectories by probe score to suppress post-commitment reasoning theater. This produces shorter, more faithful chains on GSM8K, LiveCodeBench, ToolUse, and MMLU-Redux for both Llama-8B and Qwen-7B, with gains in faithful fraction and no accuracy drop. They also release probe weights and rollouts, which helps verification.

Referee Report

3 major / 2 minor

Summary. The paper introduces ProFIL, a drop-in extension to GRPO that trains a single multi-head attention probe on frozen base-model activations (using verifier-derived labels, no human annotation) to detect post-commitment reasoning theater. During RL optimization, rollouts exceeding a probe-score threshold have their advantages zeroed. The central claim is that this yields a stable signal that suppresses theater (11-100% reduction), raises faithful CoT fraction (e.g., +24pp on LiveCodeBench via Claude judge), shortens chains (4-19%), and preserves or improves task accuracy across GSM8K, LiveCodeBench, ToolUse, and MMLU-Redux on Llama-8B and Qwen-7B, while outperforming a matched length-penalty baseline. All probe weights, configurations, and rollouts are released.

Significance. If the probe remains reliable after policy updates, ProFIL provides a practical, annotation-free route to more faithful and efficient CoT in RL-tuned reasoning models. The release of artifacts, consistent cross-domain/architecture results, and explicit comparison to length-penalty baselines are strengths that would allow the community to verify and extend the semantic-commitment filtering approach.

major comments (3)

[§3 and §4] §3 (Method) and §4 (Experiments): The central claim that the probe 'provides a stable signal that suppresses theater while resisting the RL-obfuscation failure mode' depends on the fixed probe continuing to identify post-commitment steps after GRPO updates the policy. The manuscript does not report re-measuring probe accuracy, precision, or recall on activations from the updated policy, leaving open the possibility of representation shift that could either miss theater or zero advantages on faithful trajectories.
[§4.3] §4.3 (Evaluation): The +24pp faithful-fraction gain on LiveCodeBench is measured by an independent Claude 3.7 Sonnet judge, but no details are given on the judge prompt, inter-annotator agreement, or how 'faithful' vs. 'theater' is operationalized for the judge; this makes it difficult to assess whether the reported improvement is robust to judge choice.
[Table 1 / §4.1] Table 1 / §4.1: The abstract states 'consistent gains' and 'preserving or improving task accuracy,' yet no per-domain accuracy deltas, standard errors, or statistical significance tests are referenced. Without these, it is impossible to confirm that accuracy preservation is not the result of post-hoc selection or insufficient power.

minor comments (2)

[Abstract] Abstract: The theater-reduction range '11--100%' is reported without per-domain breakdowns; adding a table row or parenthetical per-domain values would improve interpretability.
[§3.2] §3.2: The probe-score threshold is described as a hyperparameter but is not given an explicit equation or default value; including a short equation (e.g., Eq. (1)) would clarify the filtering rule.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive feedback. We address each of the major comments point by point below.

read point-by-point responses

Referee: [§3 and §4] §3 (Method) and §4 (Experiments): The central claim that the probe 'provides a stable signal that suppresses theater while resisting the RL-obfuscation failure mode' depends on the fixed probe continuing to identify post-commitment steps after GRPO updates the policy. The manuscript does not report re-measuring probe accuracy, precision, or recall on activations from the updated policy, leaving open the possibility of representation shift that could either miss theater or zero advantages on faithful trajectories.

Authors: We agree that demonstrating the probe's continued reliability on post-update activations would bolster the central claim. Although the consistent performance improvements across domains suggest the probe signal remains effective, we did not include this analysis in the original submission. In the revised manuscript, we will add results re-measuring the probe's accuracy, precision, and recall on activations sampled from the final GRPO policies using verifier labels. This will directly test for representation shift and confirm resistance to obfuscation. revision: yes
Referee: [§4.3] §4.3 (Evaluation): The +24pp faithful-fraction gain on LiveCodeBench is measured by an independent Claude 3.7 Sonnet judge, but no details are given on the judge prompt, inter-annotator agreement, or how 'faithful' vs. 'theater' is operationalized for the judge; this makes it difficult to assess whether the reported improvement is robust to judge choice.

Authors: We will include the complete judge prompt and the precise operationalization of faithfulness (post-commitment steps unrelated to the final answer, as defined in §2) in an appendix. As the evaluation uses a single LLM judge rather than multiple human annotators, traditional inter-annotator agreement does not apply; however, we will add a validation study reporting agreement between the Claude judge and human labels on a random subset of 100 examples from LiveCodeBench to support robustness. revision: yes
Referee: [Table 1 / §4.1] Table 1 / §4.1: The abstract states 'consistent gains' and 'preserving or improving task accuracy,' yet no per-domain accuracy deltas, standard errors, or statistical significance tests are referenced. Without these, it is impossible to confirm that accuracy preservation is not the result of post-hoc selection or insufficient power.

Authors: Table 1 provides the absolute accuracies per domain and model, from which deltas can be computed, but we acknowledge the value of explicit reporting. In the revision, we will update Table 1 to include per-domain accuracy deltas, standard errors (computed from the multiple training runs where applicable), and indicate that all changes are statistically insignificant (p > 0.05) based on appropriate tests. This will clarify that accuracy is preserved within the variance of the experiments. revision: yes

Circularity Check

0 steps flagged

Empirical probe-filtered RL method relies on external verifier labels and released rollouts; no derivation reduces to fitted inputs by construction

full rationale

The paper introduces ProFIL as a practical extension to GRPO: a single probe is trained once on frozen base-model activations using verifier-derived labels (no human annotation), then used to zero advantages on high-score rollouts. All reported gains (theater reduction, faithful-fraction increase, chain shortening) are measured end-to-end on held-out domains with independent judges and released artifacts. No equations define a quantity in terms of itself, no fitted parameter is relabeled as a prediction, and no load-bearing premise rests solely on self-citation. The minor risk flagged by the reader (possible representation shift or verifier-model overlap) is an empirical assumption, not a circular reduction in the derivation chain. The method is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that internal activations contain detectable signals of commitment timing and that zeroing advantages on detected theater does not harm policy optimization; no new physical entities or unstated mathematical axioms beyond standard RL and attention mechanisms.

free parameters (1)

probe score threshold
Value above which rollout advantage is zeroed; must be chosen or tuned per domain but not quantified in abstract.

axioms (1)

domain assumption Internal activations of the frozen base model contain reliable signals for detecting post-commitment reasoning steps
Invoked when training the probe on verifier-derived labels to filter GRPO rollouts.

pith-pipeline@v0.9.0 · 5600 in / 1328 out tokens · 62042 ms · 2026-05-13T02:03:00.645779+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages · 7 internal anchors

[1]

L1: Controlling how long a reasoning model thinks with reinforcement learning

Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning. InConference on Language Modeling (COLM), 2025

work page 2025
[2]

The internal state of an LLM knows when it’s lying

Amos Azaria and Tom Mitchell. The internal state of an LLM knows when it’s lying. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 967–976, 2023

work page 2023
[3]

Discovering latent knowledge in language models without supervision

Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. Discovering latent knowledge in language models without supervision. InInternational Conference on Learning Representations (ICLR), 2023

work page 2023
[4]

Reasoning Models Don't Always Say What They Think

Yanda Chen, Joe Benton, Ansh Radhakrishnan, Jonathan Uesato, Carson Denison, John Schul- man, Arushi Somani, Peter Hase, Misha Wagner, Fabien Roger, Vlad Mikulik, Samuel R. Bowman, Jan Leike, Jared Kaplan, and Ethan Perez. Reasoning models don’t always say what they think.arXiv preprint arXiv:2505.05410, 2025

work page internal anchor Pith review arXiv 2025
[5]

Training Verifiers to Solve Math Word Problems

Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

work page internal anchor Pith review Pith/arXiv arXiv 2021
[6]

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[7]

Aryo Pradipta Gema, Joshua Ong Jun Leang, Giwon Hong, Alessio Devoto, Alberto Carlo Maria Mancino, Rohit Saxena, Xuanli He, Yu Zhao, Xiaotang Du, Mohammad Reza Ghasemi Madani, Fazl Barez, Hannaneh Hajishirzi, Stuart Yu, Mark Steedman, and Pasquale Minervini. Are we done with MMLU? InProceedings of the 2025 Conference of the Nations of the Americas Chapter...

work page 2025
[8]

RLFR: Reinforcement learning from feature rewards

Goodfire AI. RLFR: Reinforcement learning from feature rewards. Goodfire AI Research (https://www.goodfire.ai/research/rlfr), 2026

work page 2026
[9]

Rl-obfuscation: Can language models learn to evade latent-space monitors?arXiv preprint arXiv:2506.14261, 2025

Rohan Gupta and Erik Jenner. Rl-obfuscation: Can language models learn to evade latent-space monitors?arXiv preprint arXiv:2506.14261, 2025

work page arXiv 2025
[10]

Measuring massive multitask language understanding

Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. InInternational Conference on Learning Representations (ICLR), 2021. 9

work page 2021
[11]

LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code

Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Ar- mando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code.arXiv preprint arXiv:2403.07974, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[12]

Kimi k1.5: Scaling Reinforcement Learning with LLMs

Kimi Team. Kimi k1.5: Scaling reinforcement learning with llms.arXiv preprint arXiv:2501.12599, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[13]

Gonzalez, Hao Zhang, and Ion Stoica

Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with PagedAttention. InProceedings of the 29th Symposium on Operating Systems Principles (SOSP), 2023

work page 2023
[14]

Measuring Faithfulness in Chain-of-Thought Reasoning

Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, et al. Measuring faithfulness in chain-of-thought reasoning.arXiv preprint arXiv:2307.13702, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[15]

Inference- time intervention: Eliciting truthful answers from a language model

Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference- time intervention: Eliciting truthful answers from a language model. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

work page 2023
[16]

Let’s verify step by step

Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In International Conference on Learning Representations (ICLR), 2024

work page 2024
[17]

Human-level control through deep reinforcement learning.Nature, 518(7540):529–533, 2015

V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning.Nature, 518(7540):529–533, 2015

work page 2015
[18]

Steering llama 2 via contrastive activation addition

Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. Steering llama 2 via contrastive activation addition. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), pages 15504–15522, 2024

work page 2024
[19]

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[20]

Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

work page 2023
[21]

Math-shepherd: Verify and reinforce llms step-by-step without human annotations

Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), 2024

work page 2024
[22]

Chi, Quoc V

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V . Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. InAdvances in Neural Information Processing Systems (NeurIPS), 2022

work page 2022
[23]

Stop reasoning now. Based on what you have written above, give your final answer in the format \boxed{}

Edwin B Wilson. Probable inference, the law of succession, and statistical inference.Journal of the American Statistical Association, 22(158):209–212, 1927. A Extended Results A.1 Length-controlled faithfulness A.2 Frozen-Probe AUROC (Anti-Gaming Evidence) A probe-as-reward system is vulnerable to RL obfuscation [ 9]: the policy could learn to produce low...

work page 1927
[24]

Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects

Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...

work page