Reinforcement learning fine-tuning of large language models increasingly adopts multiple reward dimensions, including verifiable rules, task-specific evaluators, and learned reward models, to provide richer supervision across diverse capabilities. These dimensions are commonly scalarized with fixed aggregation weights. We identify a failure mode in which aggregation itself induces reward hacking: static projection aliases qualitatively different reward profiles into a single scalar, steering optimization toward whichever dimensions are easiest, densest, or systematically favored by the reward signal. Over training, this traps the policy in suboptimal profiles and prevents convergence to better-balanced ones that would yield higher task performance. To address this, we propose Adaptive Multi-Reward Projection (AMRP), a lightweight online method that reallocates aggregation weights using three signals, relative shortfall, reward volatility, and recent progress, increasing pressure on lagging, unstable, or stagnant dimensions while relieving saturated ones. Across structured reasoning, citation-grounded generation, and open-ended alignment under GRPO, AMRP consistently improves reward-profile balance and downstream performance over fixed and dynamic weighting baselines; it also remains effective with GDPO and PPO, supporting compatibility across RL algorithms. Our code is available at https://github.com/yyhappier/AMRP.git.
To mitigate the time complexity of generative models, one-step generative models have recently emerged through direct mapping from noise to data in a single forward pass. However, the reward-guided fine-tuning method of one-step generative models remains largely unexplored. To address this, we consider one-step generators from an optimal transport view, investigating Wasserstein Gradient Flow (WGF) for modeling smooth and controlled distributional evolution in probability space. We then propose a novel reward-guided fine-tuning of a one-step generative model via WGF. We derive a practical training method that requires no reward gradients, thereby handling both non-differentiable and differentiable rewards. Moreover, our method provides smooth and stable reward-guided distributional updates while mitigating reward hacking and mode collapse. Experiments on 2D synthetic data, CIFAR-10, and ImageNet 256$\times$256 with diverse rewards, including JPEG (in)compressibility, class probability, Black-and-White and CLIP alignment, show that our method achieves better reward alignment compared to baselines.
Preference optimization is a standard alignment method for generative models, yet extending it to continuous-time dynamics remains non-trivial. In flow matching, reward-driven updates modify transport trajectories without an inherent constraint to the pretrained data manifold and can move terminal samples off the pretrained support. We formalize this failure mode as manifold drift. Theoretically, we show that optimal flow matching recovers the terminal data distribution, whereas a preference update leaves the pretrained manifold whenever its induced terminal displacement has a nonzero normal component. As a remedy, we propose ThermoDPO, a temperature-controlled objective that anchors pairwise preference optimization on preferred samples. Across temperature regimes, this objective connects rejection sampling fine-tuning and FlowDPO and controls a pointwise reconstruction-based surrogate for manifold distance. To counteract diminished signals at low temperatures, we further introduce a weighted variant, ThermoDPO-weighted. On the main toy benchmark, ThermoDPO-weighted attains a StrictScore of 0.899, compared with 0.629 for FlowDPO and 0.857 for FlowDPO+RFT. On SD3.5-M at CFG = 4.5, it improves OCR by 47.5% and the average of four metrics by 16.0%.
Reinforcement learning against rubrics, lists of criteria graded by an LLM judge, has become a standard way to post-train language models on tasks with no deterministic answer. The rubric, however, is a fixed proxy for quality, never a complete description of it, and a policy trained against it long enough will learn to exploit the difference. We measure this directly. Training Qwen3-8B with Group Relative Policy Optimization (GRPO) on medical and science rubrics and grading out-of-distribution (OOD) benchmarks with both the training judge and a stronger gold judge, we find that the two scores diverge during training. The training judge's score keeps climbing while the gold judge's score peaks and then falls, by 3 points on HealthBench-Hard and by 22 points on ResearchQA. A judge with a fixed bias would shift the gold curve by a constant, not send it down while the training score rises, so the divergence is reward hacking, not judge noise. We propose Rubric Dropout, a one-line fix borrowed from neuron dropout. At every step, we randomly drop a subset of the rubric's criteria before computing the reward, so the policy never optimizes the same rubric twice. The dropped subset is shared across each rollout group, so GRPO's group-relative advantages stay comparable, and evaluation always uses the full rubric. Comparing no dropout against dropout at 30% and 50% on both benchmark pairs, dropout raises the OOD gold score at every matched checkpoint (+1 to +2 points on HealthBench-Hard, +6 to +7 points on ResearchQA), lowers the two hacking measures we track, and costs nothing in domain. Sweeping the dropout fraction shows a broad 30-50% sweet spot, while the natural alternative, reweighting criteria by how useful they are to training, performs worse than no intervention at all in our setting.
Reinforcement learning (RL) on open-ended tasks compresses an LLM's rubric-based evaluation into a scalar reward, discarding rich textual feedback and conflating responses with distinct quality profiles. We propose Experiential Learning (EL), which repurposes the feedback model from an LLM-as-a-Judge into an LLM-as-a-Coach. The coach distills its assessment of each on-policy response into transferable experiential knowledge, which conditions a teacher model and is internalized by the policy through on-policy context distillation. Compared with scalar rewards, this higher-bandwidth feedback channel provides dense supervision and preserves fine-grained preferences among high-quality responses. Across two policy families, with feedback from the policy itself or a proprietary model, EL consistently outperforms rubric-based RL on held-out and unseen open-ended tasks. Notably, EL generalizes better beyond the training distribution, and mitigates reward hacking. These findings establish experiential knowledge as a richer and more generalizable learning signal for post-training on non-verifiable tasks.
The test suites used as RLVR rewards for code have natural false positives: per-task, persistent, asymmetric errors that accept the same wrong programs every time they appear, unlike the symmetric or resampled noise assumed by existing noise-robustness analyses. We run a preregistered two-arm causal contrast on a deployed suite: GRPO on identical MBPP tasks, seeds, and compute, rewarded by the original MBPP tests (leaky) versus the MBPP+ extra tests (hardened). Two further families replicate the design under a preregistration frozen before their data existed. [C] The average held-out effect is bounded: non-inferior under a preregistered 1.5-pt margin (gap 0.20 pt, one-sided 95% upper bound 0.75 pt). [C] Rewarded false-positive mass tracks a cheap static leakiness audit computed before training (Spearman 0.80), and the registered train-side test puts the leak-stratum FP share +43.8 pt above clean tasks. [E] Auditing every rewarded FP under signed, human-adjudicated rules finds a large residual of verified genuinely wrong code: 47.57% record-weighted; both replication families reproduce a large share. The reward paid for real bugs, not merely suite artifacts. [E] Mechanism evidence is consistent with selection of pre-existing error modes rather than learned exploitation: FP incidence does not grow within our horizon, and untrained base models already produce the same wrong outputs under the leaky filter. We then turn the same instrument on the frontier judges themselves: on their own false positives they self-assess only weakly, a same-author test is unresolved, and even the highest-scoring reader we probe stays far below its score on a weaker policy's errors -- two subjects on MBPP, licensing nothing about frontier models in general. A cheap static audit locates exposure before training; hardening the reward removes the measurement inflation, though here it buys little capability.
Does a reinforcement-learning agent that earns high reward actually learn its task's hidden state, or only a shortcut that correlates with reward? We build an instrument that makes this question directly measurable: the task is a hidden finite automaton that the agent partially controls. Because the automaton is known, the best achievable return is computable, and raw reward becomes a normalized score. The true state is also known at every step, so a linear probe can test whether the agent tracks a state it never observes. Measured separately, reward and state learning come apart: weak on-policy RL earns reward while the state probe stays at chance. Whether the agent escapes this shortcut depends on the optimizer, the training budget, and the task's structure. Structure gives an early warning: when the automaton is a permutation (group-language) automaton, a property readable from its transition table before any training, the agent will likely fail to learn the state. On a stratified set of 153 held-out automata the warning is right in 89 of the 103 cases where it fires; its absence, though, is no guarantee of safety. The instrument also tells apart two failures that reward alone cannot: a perception gap, where the linear probe finds no trace of a state the network could have represented, and a planning gap, where the state is decodable yet reward stays low. High reward alone is not evidence that the task's state was learned.
In this paper, we consider the setting where large language models (LLMs) are trained using reinforcement learning (RL) to simultaneously improve reasoning accuracy and verbalize its confidence. Our reward scheme uses two functions for rewarding confidence verbalized by the LLM: one when the LLM is correct and a different one when the LLM is incorrect. With a poorly designed reward scheme, the LLM may be incentivized to answer incorrectly so that it can be confident that its answer is indeed incorrect, a phenomenon that we call confidence reward hacking. We propose the concept of non-hackable confidence reward schemes and define a spectrum of such reward schemes for RL confidence calibration training in LLMs. We demonstrate that selective confidence reward hacking can occur in practical datasets with reward schemes that are not designed to be non-hackable. We also demonstrate that the reward scheme with the best calibration to accuracy tradeoff depends on the dataset and the application, and propose using the reward scheme as a hyperparameter to optimize the tradeoffs in accordance to what is important for the application. The code of our experiments is available in https://anonymous.4open.science/r/rl-confidence-calibration-9ED4/README.md.
Subramanyam Sahoo, Aman Chadha, Vinija Jain +1cs.LG cs.AI stat.ML
Conservative offline training is widely advocated as a safe foundation for subsequent online adaptation: if a policy stays close to well-supported behaviour, the argument goes, it is less likely to exploit imperfections in a learned reward model. We challenge this intuition empirically and mechanistically. We train a Qwen3-14B policy under Direct Preference Optimisation (DPO) with three levels of conservatism ($β\in \{β_{\mathrm{lo}}, β_{\mathrm{mid}}, β_{\mathrm{hi}}\}$ derived from empirical log-ratio percentiles), then adapt each checkpoint online against a learned reward ensemble (3\,$\times$\,Qwen3-1.7B) while measuring true performance on GSM8K exact-answer accuracy. We find that \emph{higher offline conservatism monotonically increases reward-hacking damage}, measured by the Goodhart gap and its area under the curve (AUGC), with Spearman $ρ= 1.0$ across all three conditions. Mechanistic analysis reveals a three-link causal chain: (i) high-$β$ DPO compresses policy entropy, (ii) Low-entropy policies generate responses with reduced diversity, concentrating in a narrow region of the reward model's training distribution (lower pairwise cosine distance), and (iii) despite this proximity, ensemble disagreement (epistemic uncertainty) increases with $β$ and is exploited faster during online optimisation. We further fit a power-law curve to the $(β, \augc)$ data and identify a practical optimal conservatism level $β^{\star}$ that balances alignment fidelity against hacking vulnerability. Our results suggest that the field needs \emph{calibrated}, not \emph{maximal}, conservatism.
Evgenii Opryshko, Umangi Jain, Igor Gilitschenskics.LG cs.AI
Reinforcement learning agents can exploit misspecified reward signals to achieve high apparent returns while failing on the intended objective, a failure mode known as reward hacking. Existing practical defenses typically constrain policy updates to stay near a known safe reference, creating a tension between suppressing hacking and permitting legitimate improvement. We propose Modification-Considering Value Learning (MCVL), which operationalizes the theoretical idea of current utility optimization for standard value-based RL. MCVL wraps an off-policy learner and treats each incoming transition as a candidate modification: it forecasts two training paths, one that includes the transition and one that does not, and scores both with a frozen bootstrapped-return estimator derived from a learned reward model and value function. The transition is admitted only if inclusion does not decrease the score. We formalize conditions under which this filtering is both safe and permissive, and instantiate MCVL with DDQN and TD3. Across four safety-relevant gridworlds and three modified MuJoCo continuous-control tasks with diverse hacking mechanisms, MCVL mitigates reward hacking while continuing to improve the intended objective. Project website: ktolnos.github.io/mcvl/.
Job-search platforms rely on low-bandwidth query interfaces that often fail to capture the high-dimensional complexity of candidate profiles. We present an end-to-end RLAIF (Reinforcement Learning from AI Feedback) framework to generate \emph{portable} job search queries, terms that abstract away seeker-specific identifiers while preserving generalizable qualifications. This task introduces a highly adversarial reward surface where policy optimization frequently exploits flaws in LLM-as-judge rubrics, resulting in degenerate verbatim-copying behaviors. We conducted comprehensive empirical experiments to isolate the impact of optimization mechanics against structured reward engineering. Our results demonstrate that for critic-free optimizers, performance is overwhelmingly dictated by robust reward shaping, rendering the specific choice of algorithm largely immaterial. While critic-free per-rollout baseline methods (RLOO and REINFORCE++) natively resist reward-hacking, the group-relative advantage normalization in GRPO appears uniquely sensitive to spurious reward signals, making it disproportionately susceptible to exploitation. We show that introducing a deterministic, rule-based reward floor to correct for rewards assigned to verbatim copying mitigates this failure mode, resulting in a substantial $+0.147$ quality improvement on a cross-family evaluation judge. Ultimately, we show that the training-time reward model inflates performance gains by $2.4\times$, confirming that the training success is fundamentally dependent on enforcing reward-shaping disciplines rather than selecting alternative optimizers.
Despite their widespread use, the role of reward models in shaping reinforcement learning is poorly understood. Reward models offer a tempting promise: they automatically estimate response quality in the absence of verifiers or human judges. Unlike "verifiable rewards" which typically produce binary scores, reward models typically produce continuous scores, allowing them to be sensitive to fine-grained differences in responses. However, we show this apparent strength is a serious weakness: many popular reward models are oversensitive, assigning different scores to equally good responses. Theoretically, we show that seemingly perfect reward models can be highly oversensitive; empirically, this oversensitivity can lead to bad policies. In place of existing notions of "reward model accuracy," we propose evaluating reward models using distinct measures of "discriminative ability" and "specificity" (the complement of oversensitivity). As a solution, we describe a training-free algorithm that uses Monte Carlo dropout on any neural reward model to produce discrete reward clusters. Theoretically, we prove there exist discretizations that reduce oversensitivity at minimal expense of discriminative ability; empirically we show, in both controlled and natural RL settings, that discretizing rewards leads to less reward hacking and better policies than training on the original rewards.
While RL has become a promising tool for refining world models, existing methods largely rely on conservative rollouts near the training distribution, limiting exploration, behavioral diversity, and richer dynamic discovery. In this work, we challenge this conservative paradigm. We argue that the core limitation is not exploration itself, but the lack of reliable verification strategies to support broader exploration. Without reliable verification, expanded exploration becomes highly susceptible to reward hacking, where policies exploit imperfect rewards without achieving genuine improvement. To evaluate this motivation, we instantiate our method in embodied world models, where physical plausibility, and task completion provide a rigorous testbed for scalable RL under complex dynamics. On the verification side, we introduce Reward as an Agent, an agentic reward framework that actively evaluates generated behaviors to provide robust reward signals and mitigate reward hacking under distribution shifts. On the exploration side, we introduce Dynamic-Aware Rollout Diversification through DynDiff-GRPO, which explicitly expands action-space exploration to diversify trajectories, broaden state-action coverage, and encourage richer embodied behaviors beyond conservative rollout regimes. By unifying Reward as an Agent with DynDiff-GRPO, we enable RL on a more reliable reward foundation with substantially diversified sampling, effectively mitigating reward hacking while yielding significant accuracy gains across multiple open-source world models, thereby demonstrating that broader exploration can scale successfully when grounded in robust verification.
Ely Hahami, Yoel Zimmermann, Ray Zhou +1cs.LG cs.AI cs.CL
Reinforcement learning from human feedback (RLHF) is bottlenecked by \emph{reward hacking}, where the policy exploits errors in a proxy reward model (RM) and produces high RM scores without genuine quality gains. A natural mitigation is \emph{pessimism}: penalizing rewards in regions where the RM is uncertain. However, standard scalar RMs provide no principled notion of uncertainty. We argue that the right object is a \emph{distributional} reward model $p(r\mid x,y)$. Under either a Bayesian inference or a KL-distributionally robust optimization (KL-DRO) lens, the KL-regularized RLHF objective admits a closed-form effective reward $\tilde r(x,y) = \pmβ\log\mathbb{E}_p[e^{\pm r/β}]$. The pessimistic branch unifies the prior heuristics for RM ensemble aggregation: mean aggregation, worst-case optimization (WCO), and uncertainty-weighted optimization (UWO) all emerge as limits or truncations of this single expression. This also clarifies the implicit assumptions of each existing rule.
Giorgio Maria Cavallazzi, Miguel Pérez-Cuadrado, Alfredo Pinelliphysics.flu-dyn cs.LG
Reinforcement-learning controllers optimise specified rewards, but in physical systems those rewards often capture only part of the true control objective. Three mechanisms through which this mismatch can produce apparent success without physical improvement are identified: incomplete accounting that omits relevant costs, constraint enforcement outside the policy that corrupts credit assignment, and observations that fail to resolve the relevant dynamics. All three are demonstrated in active drag reduction of wall-bounded turbulence, where the conservation constraint and full energy budget can be measured directly. A memoryless learnt policy reports drag reduction while raising total dissipation, collapsing to non-physical flow configurations. A recurrent multi-agent controller with the zero-mean projection embedded in the actor, temporal memory matched to the relevant timescales, and an actuation cost that bounds the wall power delivers a physically consistent control. Progress in physical reinforcement learning requires the reward, constraints, observations and evaluation metrics to represent unequivocally the physical objective.
Reinforcement learning (RL) systems typically optimize scalar reward functions that assume precise and reliable evaluation of outcomes. However, real-world objectives--especially those derived from human preferences--are often uncertain, context-dependent, and internally inconsistent. This mismatch can lead to alignment failures such as reward hacking, over-optimization, and overconfident behavior. We introduce a dual-source uncertainty-aware reward framework that explicitly models both epistemic uncertainty in value estimation and uncertainty in human preferences. Model uncertainty is captured via ensemble disagreement over value predictions, while preference uncertainty is derived from variability in reward annotations. We combine these signals through a confidence-adjusted Reliability Filter that adaptively modulates action selection, encouraging a balance between exploitation and caution. Empirical results across multiple discrete grid configurations (6x6, 8x8, 10x10) and high-dimensional continuous control environments (Hopper-v4, Walker2d-v4) demonstrate that our approach yields more stable training dynamics and reduces exploitative behaviors under reward ambiguity, achieving a 93.7% reduction in reward-hacking behavior as measured by trap visitation frequency. We demonstrate statistical significance of these improvements and robustness under up to 30% supervisory noise, albeit with a trade-off in peak observed reward compared to unconstrained baselines. By treating uncertainty as a first-class component of the reward signal, this work offers a principled approach toward more reliable and aligned reinforcement learning systems.