Pith. sign in

REVIEW 2 major objections 3 minor 28 references

Self-improving harness optimizers can hallucinate failures — and build guardrails for rules that provably don't exist.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 06:55 UTC pith:56KNJULY

load-bearing objection A clean, oracle-certified demonstration that a harness optimizer invents failures that never happened; the core 15/60 result holds up, but the 'three conditions' framing is cleaner than the controls actually support. the 2 major comments →

arxiv 2607.13083 v1 pith:56KNJULY submitted 2026-07-13 cs.CR cs.SE

Phantom Guardrails: When Self-Improving Agent Harnesses Fix Failures That Never Happened

Su Wang , Pin Qian , Yifan Lin , Jingzhou Xu , Yihang Chen , Xiaochong Jiang , Lifei Liu , Haoran Yu This is my paper
classification cs.CR cs.SE
keywords agentic AI evaluationharness optimizationfailure fabricationphantom guardrailoracle-based evaluationLLM agentsself-improving agentssuppression-only acceptance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper shows that automated harness optimizers — loops that revise an AI agent's scaffolding to eliminate observed failures — can fabricate failures that never occurred. In a deterministic micro-lab where every move is tagged legal and a byte-exact oracle certifies ground truth, the optimizer enables a guard for a rule the oracle proves does not exist, citing a violation the oracle refutes, in 25 percent of runs. The fabrication is not indiscriminate: it appears only when a benign pattern resembles a familiar game rule, the rule set's completeness is unstated, and the instructions presuppose failures. Because the invented guard changes no true outcome and cannot improve an already-perfect suppression score, it is invisible to acceptance rules that only check suppression, yet an add-only accept loop keeps it permanently. The paper presents the lab as an instrument for measuring such invented failures.

Core claim

The central claim is that a suppression-rewarded harness proposer, given evidence in which every move is legal, will nonetheless — under three coinciding conditions — invent a failure: it enables a special-rule guard (g_castle) for a rule that provably never occurs in the data and cites a specific violation that the oracle refutes. The same guard fires correctly when a real violation is injected, and the proposer abstains on featureless all-legal input, so the invention is a genuine false positive against ground truth rather than over-building or reward hacking. The paper identifies the mechanism as a genre-prior import: the repeated-move pattern matches a familiar board-game rule, and the p

What carries the argument

The Counterfactual Fabrication Lab: a deterministic micro-lab that plants a guard (g_castle) for a failure class the task cannot produce, presents only legal episodes, and scores proposals with a byte-exact oracle. The key mechanism is the suppression proxy — the fraction of episodes left with no firing failure class — which is already maximal on all-legal pools, so no edit can improve it and the warranted harness is empty. The fabrication metric reads whether g_castle was enabled on a pool the oracle certifies free of that class. The load-bearing comparison is the three-pool contrast (congruent, fabrication, pristine) plus the battery of planted patterns and control arms that isolate each o

Load-bearing premise

The load-bearing premise is that the repeated-square pattern stands in for the whole category of 'rule-shaped patterns resembling a familiar game rule'; the three-condition claim is demonstrated for this one pattern, and other genre rules (capture, check, turn-taking) were not tested, so the mechanism's breadth is unmeasured.

What would settle it

Run the lab's fabrication pool with additional rule-shaped patterns from the same board-game genre (e.g., a move sequence resembling a capture or a check) at the same fixed incidence; if the phantom guard is enabled at a rate statistically indistinguishable from zero for several such patterns while the repeated square reproduces, the 'rule-shaped pattern' condition is specific to repetition rather than general, and the three-condition mechanism does not extend to the category.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Suppression-only acceptance rules are blind to fabricated no-op guards: a guard that cannot change the suppression score is never demerited, so add-only loops accumulate phantoms and never remove them.
  • Instruction hygiene (not presupposing failures) and specification completeness (stating the rule set is complete) each drive the single-shot fabrication rate from 15/60 to 0/60.
  • Warrant-aware acceptance — crediting a guard only when the proposer cites an episode whose failure the oracle confirms the guard suppresses — excludes the phantom entirely while still adopting the real fixers.
  • The effect is not reward hacking (no true-return loss, no proxy gain) and not over-refusal (no helpfulness trade-off); it is a distinct failure mode: a fix for a failure that never happened.
  • In an add-only accept loop, the phantom accumulates monotonically even under a neutral charter and a displayed perfect suppression score; the measured per-round entry rate is 0.050.

Where Pith is reading between the lines

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

  • If the mechanism generalizes, deployed self-improving agents that optimize their own scaffolds from post-hoc failure signals may silently accumulate unused guardrails, increasing latency and attack surface without improving outcomes — a tax invisible to suppression-based benchmarks.
  • The three-condition gate suggests a testable prediction: any task domain with a strong genre prior (e.g., security rules around tool use) should show the same fabrication when that prior's rule shape appears benignly, provided the charter presupposes failures and the taxonomy is uncertified.
  • The add-only result implies that merely removing failure presupposition from prompts is insufficient in iterative loops; acceptance rules must check warrant, not just suppression, to keep phantom guards out.
  • The abstention on featureless input indicates the proposer is not a compulsive builder; mapping the boundary of the 'rule-shaped pattern' category across a broader set of prior-driven regularities would sharpen the mechanism's scope.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 3 minor

Summary. The paper introduces the Counterfactual Fabrication Lab, a deterministic micro-benchmark in which an LLM-based harness proposer edits a fixed set of guard hooks and is scored by a suppression proxy. On an all-legal pool containing a benign repeated-square regularity, the proposer enables a guard for a nonexistent 'special rule' in 15/60 runs, versus 0/60 on a featureless pool (z=4.14). The authors argue the effect is structured: it requires a rule-shaped pattern, an unstated rule-set completeness, and a failure-presupposing instruction, and they support this with battery, completeness-assured, and neutral-instruction arms, plus a congruent-pool control and a blinded lexical control for genuine violations. They further show that in an add-only accept loop, suppression-only acceptance ratchets the phantom guard in (11/60 by round 4), while warrant-aware acceptance excludes it (0/60). The paper is unusually careful about run accounting, oracle-based ground truth, and verbatim rationales.

Significance. If the central claim holds, the paper identifies a failure mode that is distinct from reward hacking and over-refusal: a suppression-rewarded optimizer can invent a failure class that provably never occurs and harden against it, with no effect on the reward signal. The lab is a genuinely useful instrument because the oracle is byte-exact and the warranted action is known by construction, making the false positive auditable rather than model-judged. The inclusion of per-run raw rationales, deterministic audit gates, and reproducible rebuilds ('make review') substantially raises confidence. The main result is statistically modest and concentrated in one proposer, but the authors acknowledge this and frame the contribution as categorical and mechanistic. The principal weakness is that one putative necessary condition — the absence of the 'special-rule' wording on hook S — is never varied on an all-legal, pattern-bearing pool, so the 'three conditions' decomposition is not fully identified.

major comments (2)
  1. [§3.2, §4.3, §4.4] The claim that fabrication appears 'only when three conditions coincide' is not fully identified, because the hook-S description 'may block a special-rule violation' is never removed on an all-legal pool that carries the repeated-square pattern. §3.2 flags this lexical frame, but the controls do not rule it out as a necessary cause: the §4.3 battery holds S's wording fixed and varies only the pattern, showing the pattern is necessary in the presence of the label, not that the label is dispensable; the §4.4 blinded-congruent arm removes the label but runs on a pool with a genuine illegal move, establishing routing rather than abstention. Thus a fourth condition — the generic 'special-rule' hook description — is confounded with the pattern in every fabrication cell. Please add a control: run the fabrication pool (all-legal repeated square) with S described by elimination exactly as in the
  2. [§4.2, Appendix C] The phrase 'oracle refutes every cited violation' overstates what O_castle checks. O_castle fires only on CASTLE∧ILLEGAL records; the fabrication rationales cite 'repeated moves to the same position' or 'already-occupied squares,' which are not among the three oracle classes. The assertion that no such rule exists is true by the lab's completeness assumption, not by the byte-exact oracle itself. The oracle certifies only that no castle-class violation occurred in the pool. Please rephrase to say the oracle refutes the cited castle-class violation, or explicitly note that the broader refutation relies on the task-level completeness assumption. This matters because the abstract's 'byte-exact oracle to check every cited violation' invites readers to think the oracle checks the specific invented rule.
minor comments (3)
  1. [§4.5] The '2×2 of §4.3' is presented as an inline table without a number; consider adding a formal table number for cross-referencing. Also, the 'n.+a.' column in Table 4 is not defined in the caption; please spell out 'neutral instruction + completeness-assured'.
  2. [§4.2] The extension roster is described as post-hoc, and the paper does not apply a multiple-comparison correction across the many arms and per-proposer cells. Given the acknowledged concentration of the effect in one proposer, a brief statement about the exploratory nature of the per-model comparisons would strengthen the reporting.
  3. [§3.1] Equation (3) defines fabrication as enabling g_castle on a pool with O_castle≡0. The paper should be consistent in wording: the 'fabricated rule' in the rationales is not literally 'castle,' but a generic 'special rule.' Clarifying the mapping between the menu label and O_castle would avoid the ambiguity raised in the second major comment.

Circularity Check

0 steps flagged

No significant circularity: the phantom-guard result is an externally oracle-grounded measurement, not a fitted prediction, and no self-citation chain is load-bearing.

full rationale

The paper's central chain—an oracle-certified phantom class, the fabrication metric Eq. (3), measurement on all-legal pools, and the §4.3 controls—is not circular. The phantom class O_castle is certified by a deterministic oracle reading disjoint record fields, independently of the proposer's behavior. Eq. (3) fixes 'fabrication' before any run, and the headline rates (15/60 vs. 0/60, z=4.14) are empirical counts with confidence intervals, not identities derived from the definition. No parameter is fitted to data and then renamed as a prediction; the per-round entry rate q is measured, and the accept-loop dynamics are verified against a deterministic judge and live proposers. The §4.3 'three conditions' claim is supported by held-fixed manipulations of pattern shape, completeness assurance, and instruction wording; the paper also acknowledges the single-genre scope and the menu's 'special-rule' lexical frame (§3.2, §5), which are genuine experimental limitations or confound risks, not circularity. There are no load-bearing self-citations: none of the cited references are by the present authors, and external works are used for contrast (OR-Bench, Skalse et al., RHO), not as the source of the mechanism. The only definitional proximity is that enabling g_castle on an oracle-clean pool is by definition 'fabrication,' but the classification is grounded in external ground truth, and the observed rationales—citing moves tagged legal as violations—go beyond the metric definition. One wording overstatement ('cites a violation the oracle refutes') should be read as refuted by the generator's completeness assumption rather than by O_castle directly, but this is a precision issue, not circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 1 invented entities

The central claims rest on accepting the oracle's taxonomy as complete and the lab as representative of deployed harness search; the category-level generalization from one rule-shaped pattern is the largest unmeasured step. No free parameters are fitted to force the result.

axioms (5)
  • domain assumption The legality tags and the failure taxonomy C={syntax,bounds,castle} are the complete and authoritative definition of what counts as a failure in MiniArena.
    §3.1 fixes the oracle to these classes; the 'fabrication' label is meaningful only if rules outside C (e.g., repetition) truly do not exist, which is true by construction of the generator but is an assumption the reader must accept.
  • domain assumption LLM API proposers at temperature 0.7 with seeds are a faithful proxy for deployed self-improving harness optimizers.
    The lab uses five/eight commercial models and calls this 'the regime we study' (§1, §3.2); no deployed-harness evidence is provided, so generalization rests on representativeness.
  • ad hoc to paper The repeated-square regularity is representative of the category 'rule-shaped pattern resembling a familiar game rule'.
    §4.3 tests one rule-shaped pattern (repeated square) and three non-rule patterns; the three-condition claim uses the category-level phrase 'rule-shaped pattern,' but sufficiency for other rule-shaped patterns is untested.
  • domain assumption The add-only accept loop is a faithful model of suppression-rewarded maintenance loops.
    §4.5 calls this 'a deliberate modeling choice'; persistence is guaranteed by the never-subtract construction, so the loop result is partly by design rather than fully emergent.
  • domain assumption The suppression proxy S(H,D) is the only reward signal in the studied regime.
    Eq. (1) defines the regime; this is an explicit modeling choice, and the invisibility conclusion follows from it.
invented entities (1)
  • Phantom guardrail independent evidence
    purpose: Names the phenomenon where a harness proposer adds a guard for a failure class certified by oracle to never occur; it is the object of study and the fabrication metric Fab(H,D) in Eq. (3).
    Operationalized as g_castle enablement on oracle-certified clean pools (15/60 vs 0/60), so the construct has a falsifiable handle outside the narrative; it is a conceptual label, not a physical entity.

pith-pipeline@v1.3.0-alltime-deepseek · 21228 in / 15915 out tokens · 152182 ms · 2026-08-02T06:55:23.610065+00:00 · methodology

0 comments
read the original abstract

Self-improving AI agents are designed to learn from their mistakes. We show they can also hallucinate mistakes that never happened. We study this failure mode in automated harness optimization, where an LLM-based proposer edits an agent's scaffold, including prompts, parsers, filters, validators and guardrails, to eliminate observed failures. But this process rarely asks first: was there a real failure to fix? We introduce the Counterfactual Fabrication Lab, a deterministic micro-lab where the correct action is known: do nothing. The lab plants a candidate guardrail for a failure class that provably never occurs, presents only legal episodes, and uses a byte-exact oracle to check every cited violation. The proposer behaves as expected on real violations and abstains on featureless legal input. Yet when the legal input contains a harmless pattern resembling a familiar game rule, it invents a failure: in 15/60 runs, versus 0/60 on featureless input, it enables the nonexistent-rule guardrail and cites a violation the oracle refutes. The effect is structured, not indiscriminate. In single-shot proposals it appears only when three conditions coincide: a rule-shaped pattern, an open-ended rule set and an instruction that presupposes failures. Removing any of these conditions eliminates the fabrication. Because the invented guardrail changes no true outcome and cannot improve an already-perfect suppression score, the phenomenon is neither reward hacking nor over-refusal. It is a phantom guardrail: a fix for a failure that never happened, invisible to suppression-only acceptance. Inside an add-only accept loop it re-enters even without the failure-presupposing instruction, the loop's keep-adding role supplying the demand the instruction supplied in single shot, and once in it stays. We present the Counterfactual Fabrication Lab for measuring fabricated failures in self-improving agent harnesses.

Figures

Figures reproduced from arXiv: 2607.13083 by Haoran Yu, Jingzhou Xu, Lifei Liu, Pin Qian, Su Wang, Xiaochong Jiang, Yifan Lin, Yihang Chen.

Figure 1
Figure 1. Figure 1: Oracle-checked fabrication. (Left) The same phan￾tom guard g_castle: enabled 1.00 when a real violation is present (the detector works, all five proposers), 0.00 on fea￾tureless input (abstention), but 0.25 on an all-legal pool, a false positive for a rule the oracle proves does not exist (fab￾rication vs. pristine, two-proportion 𝑧 = 4.14). (Right) Per￾proposer fabrication rate (five protocol and three ex… view at source ↗
Figure 2
Figure 2. Figure 2: The mechanism controls (all pools all-legal; [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Accept-loop dynamics (phantom-in-deployed [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Matching-fixer selection rate by proposer × equivalence-regime arm (GridErrand, 10 per cell). Each cell is the rate the matching fixer was built, a separate quantity from the mass contrast Δ of [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 19 linked inside Pith

  1. [1]

    Justin Cui, Wei-Lin Chiang, Ion Stoica, and Cho-Jui Hsieh. 2024. OR-Bench: An Over-Refusal Benchmark for Large Language Models. Appeared at ICML 2025. arXiv:2405.20947 [cs.CL]

    Pith/arXiv arXiv 2024

  2. [2]

    Sunishchal Dev, Andrew Sloan, Joshua Kavner, Nicholas Kong, and Morgan Sandler. 2026. Judge Reliability Harness: Stress Testing the Reliability of LLM Judges. Accepted at Agents in the Wild: Safety, Security, and Beyond Workshop at ICLR 2026. arXiv:2603.05399 [cs.AI]

    arXiv 2026

  3. [3]

    Yuxuan Jiang and Francis Ferraro. 2026. SCRIBE: Structured Mid-Level Supervi- sion for Tool-Using Language Models.arXiv preprint arXiv:2601.03555(2026). arXiv:2601.03555 [cs.AI] doi:10.48550/arXiv.2601.03555

    Pith/arXiv arXiv doi:10.48550/arxiv.2601.03555 2026

  4. [4]

    Changxin Ke, Rui Zhang, Jiaming Guo, Yuanbo Wen, Li Ding, Shuo Wang, et al

  5. [5]

    Yoonho Lee, Roshen Nair, Qizheng Zhang, Kangwook Lee, Omar Khattab, and Chelsea Finn. 2026. Meta-Harness: End-to-End Optimization of Model Harnesses. arXiv:2603.28052 [cs.LG]

    Pith/arXiv arXiv 2026

  6. [6]

    Samuel Lewis-Lim, Xingwei Tan, Zhixue Zhao, and Nikolaos Aletras. 2025. Analysing Chain of Thought Dynamics: Active Guidance or Unfaithful Post-hoc Rationalisation? arXiv:2508.19827 [cs.CL]

    Pith/arXiv arXiv 2025

  7. [7]

    Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, Kai Shu, Lu Cheng, and Huan Liu. 2025. From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. Associ...

    doi:10.18653/v1/2025.emnlp-main.138 2025

  8. [8]

    Yanhang Li, Zhichao Fan, and Zexin Zhuang. 2026. SafetyRepro: Configuration-Conditional Rank Instability on Alignment Benchmarks. arXiv:2605.25492 [cs.LG] doi:10.48550/arXiv.2605.25492

    Pith/arXiv arXiv doi:10.48550/arxiv.2605.25492 2026

  9. [9]

    Jiahang Lin, Shichun Liu, Chengjun Pan, Lizhi Lin, Shihan Dou, Zhiheng Xi, Xuanjing Huang, Hang Yan, Zhenhua Han, Tao Gui, and Yu-Gang Jiang. 2026. Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses. arXiv:2604.25850 [cs.CL]

    Pith/arXiv arXiv 2026

  10. [10]

    Lixing Lin, Juli You, Yue Li, Luyun Lin, Yiqing Wang, Zhen Zhang, and Mox- uan Zheng. 2026. Reflect-Guard: Enhancing LLM Safeguards against Adversar- ial Prompts via Logical Self-Reflection.arXiv preprint arXiv:2605.24834(2026). arXiv:2605.24834 [cs.CR] doi:10.48550/arXiv.2605.24834

    Pith/arXiv arXiv doi:10.48550/arxiv.2605.24834 2026

  11. [11]

    Bissyandé, Dongsun Kim, Jacques Klein, and Yves Le Traon

    Kui Liu, Anil Koyuncu, Tegawendé F. Bissyandé, Dongsun Kim, Jacques Klein, and Yves Le Traon. 2018. You Cannot Fix What You Cannot Find! An Investigation of Fault Localization Bias in Benchmarking Automated Program Repair Systems. Appeared at ICST 2019. arXiv:1812.07283 [cs.SE]

    Pith/arXiv arXiv 2018

  12. [12]

    Xinghua Lou, Miguel Lázaro-Gredilla, Antoine Dedieu, Carter Wendelken, Wolf- gang Lehrach, and Kevin P. Murphy. 2026. AutoHarness: Improving LLM Agents by Automatically Synthesizing a Code Harness. arXiv:2603.03329 [cs.AI]

    arXiv 2026

  13. [13]

    David Manheim and Scott Garrabrant. 2018. Categorizing Variants of Goodhart’s Law. arXiv:1803.04585 [cs.AI]

    Pith/arXiv arXiv 2018

  14. [14]

    Xuying Ning, Katherine Tieu, Dongqi Fu, Tianxin Wei, Zihao Li, Yuanchen Bei, et al. 2026. Code as Agent Harness: Toward Executable, Verifiable, and Stateful Agent Systems. arXiv:2605.18747 [cs.AI]

    Pith/arXiv arXiv 2026

  15. [15]

    Wenbo Pan, Shujie Liu, Chin-Yew Lin, Jingying Zeng, Xianfeng Tang, Xi- angyang Zhou, Yan Lu, and Xiaohua Jia. 2026. Retrospective Harness Opti- mization: Improving LLM Agents via Self-Preference over Trajectory Rollouts. arXiv:2606.05922 [cs.CL]

    Pith/arXiv arXiv 2026

  16. [16]

    Joar Skalse, Nikolaus H. R. Howe, Dmitrii Krasheninnikov, and David Krueger

  17. [17]

    Chung-En Sun, Linbo Liu, Ge Yan, Zimo Wang, and Tsui-Wei Weng. 2026. LLM Agents Already Know When to Call Tools – Even Without Reasoning. arXiv:2605.09252 [cs.AI]

    Pith/arXiv arXiv 2026

  18. [18]

    Xian Sun, Wei Gao, Yingshuo Wang, Lingdong Kong, Yanhang Li, Zhichao Fan, Zexin Zhuang, Wenlong Dong, Zhiyuan Zheng, Hrishikesh Paranjape, Abhishek Mandal, and Johnny R. Zhang. 2026. Beyond Accuracy: Measuring Bias Acknowledgment in Chain-of-Thought Reasoning for Responsible AI Evaluation. ICML 2026 Workshop on Trustworthy AI for Good. arXiv:2606.15127 [c...

    Pith/arXiv arXiv doi:10.48550/arxiv.2606.15127 2026

  19. [19]

    Varun Ursekar, Apaar Shanker, Veronica Chatrath, Yuan Xue, and Sam Den- ton. 2026. VeRO: An Evaluation Harness for Agents to Optimize Agents. arXiv:2602.22480 [cs.AI]

    Pith/arXiv arXiv 2026

  20. [20]

    Yinghao Wang and Cheng Wang. 2026. The Observability Gap: Why Output- Level Human Feedback Fails for LLM Coding Agents. arXiv:2603.26942 [cs.SE]

    arXiv 2026

  21. [21]

    Whitson and Adam D

    Jennifer A. Whitson and Adam D. Galinsky. 2008. Lacking Control Increases Illusory Pattern Perception.Science322, 5898 (2008), 115–117. doi:10.1126/ science.1159845

    2008

  22. [22]

    Tianshi Xu, Huifeng Wen, and Meng Li. 2026. Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents. arXiv:2605.22166 [cs.AI]

    Pith/arXiv arXiv 2026

  23. [23]

    Boyang Yang, Zijian Cai, Shunfu Jin, et al . 2026. PAFT: Preservation-Aware Fine-Tuning for Minimal-Edit Program Repair. arXiv:2604.03113 [cs.SE]

    Pith/arXiv arXiv 2026

  24. [24]

    Bo Yue, Shufan Wang, Ashish Gaurav, Jian Li, Pascal Poupart, and Guiliang Liu

  25. [25]

    this matters

    Jianxiang Zang, Yongda Wei, Ruxue Bai, Shiyu Jiang, Nijia Mo, Binhong Li, Qiang Sun, and Hui Liu. 2025. Reward Auditor: Inference on Reward Modeling Suitability in Real-World Perturbed Scenarios.arXiv preprint arXiv:2512.00920 (2025). arXiv:2512.00920 [cs.CL] doi:10.48550/arXiv.2512.00920 7 Wang et al. Supplementary Material Phantom Guardrails: When Self-...

    Pith/arXiv arXiv doi:10.48550/arxiv.2512.00920 2025

  26. [2022]

    Appeared at NeurIPS 2022

    Defining and Characterizing Reward Hacking. Appeared at NeurIPS 2022. arXiv:2209.13085 [cs.LG]

    Pith/arXiv arXiv 2022

  27. [2025]

    InThe Thirteenth International Conference on Learning Represen- tations

    Understanding Constraint Inference in Safety-Critical Inverse Reinforce- ment Learning. InThe Thirteenth International Conference on Learning Represen- tations. ICLR, 50327–50354. https://proceedings.iclr.cc/paper_files/paper/2025/ hash/7db81dc967383b560798b0954d51973d-Abstract-Conference.html

    2025

  28. [2026]

    arXiv:2604.05963 [cs.SE]

    QiMeng-PRepair: Precise Code Repair via Edit-Aware Reward Optimization. arXiv:2604.05963 [cs.SE]

    Pith/arXiv arXiv