cs.LGNov 16, 2026

Optimistic Verifiers Get Hacked: Reward Hacking Under an Exact Process Oracle

Authors: Arunkumar V, Manoranjan Gandhudi, Gangadharan G. R., Arun Prakash, S. Senthilkumar

Organizations: University College of Engineering, BIT Campus, Anna University, Tiruchirappalli, India · Central University of Karnataka, Kalaburagi, India · National Institute of Technology Tiruchirappalli, India · Jawaharlal Nehru University, New Delhi, India

Abstract

Reinforcement learning from verifiable rewards inherits whatever its verifier gets wrong. Outside mathematics and code the verifier is a proxy, a rubric, a learned judge or a hand-written checker, and its distance from the true objective is never measured, so a rising reward cannot be told apart from a hack. We build a setting where the true reward is exactly computable at the process level: executable authorization policies over 124,416 contexts admit an oracle that lists every minimal sufficient reason for every decision. We compile these policies into exact process rewards, degrade the verifier along seven realistic axes (outcome-only, sufficiency, recall-only, extraction noise, coarse taxonomy, stale policy, LLM judge), and train against each while the exact verifier scores every rollout. Four phenotypes appear. Optimistic verifiers are hacked by padding: the sufficiency proxy ends at 0.940 while true faithfulness falls from 0.731 to 0.667 on 96 held-out episodes. Outcome-only verifiers saturate and stop teaching, noise dilutes, and ranking-preserving gaps do no harm. A model-free audit of the verifier, its selection error and gap sign, together with a short saturation pilot, anticipates which phenotype a run will produce. Across nine configurations spanning two frameworks, two model vendors, two adapter sizes, three precisions, two initialisations and two seeds, the phenotypes recur, sufficiency damage is positive in all 10 replication cells and beyond two standard errors in 7, and optimistic gaps remain the steeper branch. The damage magnitude does not replicate, and five pre-registered attempts to explain it are reported as they fell. We release the testbed, the training instrumentation, and 14 real public policies (Cedar, OPA, Gatekeeper) ported to executable oracles.

Explore similar work

May 12, 2026

Reward Hacking in Rubric-Based Reinforcement Learning

Reinforcement learning with verifiable rewards has enabled strong post-training gains in domains such as math and coding, though many open-ended settings rely on rubric-based rewards. We study reward hacking in rubric-based RL, where a policy is optimized against a training verifier but evaluated against a cross-family panel of three frontier judges, reducing dependence on any single evaluator. Our framework separates two sources of divergence: verifier failure, where the training verifier credits rubric criteria that reference verifiers reject, and rubric-design limitations, where even strong rubric-based verifiers favor responses that rubric-free judges rate worse overall. Across medical and science domains, weak verifiers produce large proxy-reward gains that do not transfer to the reference verifiers; exploitation grows over training and concentrates in recurring failures such as partial satisfaction of compound criteria, treating implicit content as explicit, and imprecise topical matching. Stronger verifiers substantially reduce, but do not eliminate, verifier exploitation. We also introduce a self-internalization gap, a verifier-free diagnostic based on policy log-probabilities, which tracks reference-verifier quality, detecting when the policy trained using the weak verifier stops improving. Finally, in our setting, stronger verification does not prevent reward hacking when the rubric leaves important failure modes unspecified: rubric-based verifiers prefer the RL checkpoint, while rubric-free judges prefer the base model. These disagreements coincide with gains concentrated in completeness and presence-based criteria, alongside declines in factual correctness, conciseness, relevance, and overall quality. Together, these results suggest that stronger verification reduces reward hacking, but does not by itself ensure that rubric gains correspond to broader quality gains.
Sep 28, 2026

Verifier Errors in RLVR: Reward Hacking, Limits of Feedback, and Selective Control

In reinforcement learning with verifiable rewards (RLVR), imperfect verifiers can reward incorrect responses, creating opportunities for reward hacking. Using gradient flow with a fixed verifier, we characterize the conditions under which reward rises while correctness falls. We then show that the observations available during RLVR are, in general, insufficient to detect or identify accepted errors, or to guarantee their reduction without sacrificing correct responses. To address this limit, we construct a correction using additional feedback about correctness from audits. This correction achieves selective control: at the current policy, it lowers the probability of accepted errors and raises that of correct responses, provided it outweighs the pressure toward errors from verifier reward. Experiments with log linear and neural contextual bandits and with a language model support the analysis and show that selective control under partial auditing reduces accepted errors while increasing correctness.
Oct 26, 2026

Audit Where Hacks Pay: Budgeted Strong-Verifier Audits Against Reward Hacking in GRPO

Reinforcement learning with verifiable rewards trains on cheap checks, such as unit tests, that a policy can learn to game, while a stronger check is too expensive to run on every rollout. We study how to spend a limited strong-verifier budget in GRPO, separating which weakly passing rollouts to audit from what reward a caught hack receives. At a fixed policy state with centered advantages, the expected update depends on a design only through each rollout's marginal audit probability and correction; reflecting a caught hack's reward about the failure reward negates the advantages of an all-hack group; and auditing whole groups with the fewest weak passes first (low-pass budgeted audits, LB) minimizes the advantage mass left on hacks among pass-count designs with a fixed correction when a label's value falls with the pass count. We tested this analysis in actual LLM training: from each run's starting state, the sign of this mass on a frozen panel predicted whether hacking would be suppressed in 33 of 39 runs, including 13 of 16 preregistered ones. In LeetCode (Wong et al., 2026) and Countdown-Code (Khalifa et al., 2026), LB drives hacking to near zero, while random audits with the same penalty and budget do not and use more labels. We also find that the randomness of the update affects training: in Countdown-Code, two unbiased corrections, per-item reward-level IPW and a group-level advantage correction, bring hacking down too but end 6.1-11.8 and 8.0-11.5 points lower in held-out accuracy than LB at the same budget. A full audit, which the group-level correction reproduces in expectation, stays within about two points of LB in three seeds, while LB uses 28% of its labels, and allocating the same correction by pass count recovers most of the loss (one seed). Training outcomes thus depend not only on the expected update but also on the randomness of its estimate.