The Policy Cliff: A Theoretical Analysis of Reward-Policy Maps in Large Language Models

• URL: http://arxiv.org/abs/2507.20150v1
• Date: Sun, 27 Jul 2025 06:56:10 GMT
• Title: The Policy Cliff: A Theoretical Analysis of Reward-Policy Maps in Large Language Models
• Authors: Xingcheng Xu,
• Abstract summary: Reinforcement learning (RL) plays a crucial role in shaping the behavior of large language and reasoning models (LLMs/LRMs)<n>However, it often produces brittle and unstable policies, leading to critical failures such as spurious reasoning, deceptive alignment, and instruction disobedience.<n>This paper presents a rigorous mathematical framework for analyzing the stability of the mapping from a reward function to the optimal policy.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Reinforcement learning (RL) plays a crucial role in shaping the behavior of large language and reasoning models (LLMs/LRMs). However, it often produces brittle and unstable policies, leading to critical failures such as spurious reasoning, deceptive alignment, and instruction disobedience that undermine the trustworthiness and safety of LLMs/LRMs. Currently, these issues lack a unified theoretical explanation and are typically addressed using ad-hoc heuristics. This paper presents a rigorous mathematical framework for analyzing the stability of the mapping from a reward function to the optimal policy. We show that policy brittleness often stems from non-unique optimal actions, a common occurrence when multiple valid traces exist in a reasoning task. This theoretical lens provides a unified explanation for a range of seemingly disparate failures, reframing them as rational outcomes of optimizing rewards that may be incomplete or noisy, especially in the presence of action degeneracy. We extend this analysis from the fundamental single-reward setting to the more realistic multi-reward RL across diverse domains, showing how stability is governed by an "effective reward" aggregation mechanism. We also prove that entropy regularization restores policy stability at the cost of increased stochasticity. Our framework provides a unified explanation for recent empirical findings on deceptive reasoning, instruction-following trade-offs, and RLHF-induced sophistry, and is further validated through perturbation experiments in multi-reward RL. This work advances policy-stability analysis from empirical heuristics towards a principled theory, offering essential insights for designing safer and more trustworthy AI systems.

Related papers

• Learning Structured Reasoning via Tractable Trajectory Control [99.75278337895024]
  Ctrl-R is a framework for learning structured reasoning via tractable trajectory control.<n>We show that Ctrl-R enables effective exploration and internalization of previously unattainable reasoning patterns.
  arXiv  Detail & Related papers  (2026-03-02T09:18:19Z)
• Stabilizing Policy Optimization via Logits Convexity [59.242732612484474]
  We show that the convexity of the supervised fine-tuning loss with respect to model logits plays a key role in enabling stable training.<n>Motivated by this observation, we propose Logits Convex Optimization (LCO), a simple yet effective policy optimization framework.
  arXiv  Detail & Related papers  (2026-03-01T07:40:12Z)
• Towards Generalizable Reasoning: Group Causal Counterfactual Policy Optimization for LLM Reasoning [50.352417879912515]
  Large language models (LLMs) excel at complex tasks with advances in reasoning capabilities.<n>We propose Group Causal Counterfactual Policy Optimization to explicitly train LLMs to learn generalizable reasoning patterns.<n>We then construct token-level advantages from this reward and optimize the policy, encouraging LLMs to favor reasoning patterns that are process-valid and counterfactually robust.
  arXiv  Detail & Related papers  (2026-02-06T08:03:11Z)
• UCPO: Uncertainty-Aware Policy Optimization [12.847800921274617]
  Existing Large Language Models (LLMs) often suffer from Advantage Bias due to binary decision spaces and static uncertainty rewards, inducing excessive conservatism or overconfidence.<n>This paper unveils the root causes of reward hacking and overconfidence in current RL paradigms incorporating uncertainty-based rewards, based on which we propose the UnCertainty-Aware Policy Optimization framework.
  arXiv  Detail & Related papers  (2026-01-30T07:07:42Z)
• BAPO: Boundary-Aware Policy Optimization for Reliable Agentic Search [72.87861928940929]
  Boundary-Aware Policy Optimization (BAPO) is a novel RL framework designed to cultivate reliable boundary awareness without compromising accuracy.<n>BAPO introduces two key components: (i) a group-based boundary-aware reward that encourages an IDK response only when the reasoning reaches its limit, and (ii) an adaptive reward modulator that strategically suspends this reward during early exploration, preventing the model from exploiting IDK as a shortcut.
  arXiv  Detail & Related papers  (2026-01-16T07:06:58Z)
• Evidence-Augmented Policy Optimization with Reward Co-Evolution for Long-Context Reasoning [52.144281362465996]
  We propose EAPO (Evidence-Augmented Policy Optimization) to apply Reinforcement Learning to long-context scenarios.<n>We first establish the Evidence-Augmented Reasoning paradigm, validating via Tree-Structured Evidence Sampling.<n>We then introduce a specialized RL algorithm where a reward model computes a Group-Relative Evidence Reward.<n>To sustain accurate supervision throughout training, we further incorporate an Adaptive Reward-Policy Co-Evolution mechanism.
  arXiv  Detail & Related papers  (2026-01-15T11:40:57Z)
• How and Why LLMs Generalize: A Fine-Grained Analysis of LLM Reasoning from Cognitive Behaviors to Low-Level Patterns [51.02752099869218]
  Large Language Models (LLMs) display strikingly different generalization behaviors.<n>We introduce a novel benchmark that decomposes reasoning into atomic core skills.<n>We show that RL-tuned models maintain more stable behavioral profiles and resist collapse in reasoning skills, whereas SFT models exhibit sharper drift and overfit to surface patterns.
  arXiv  Detail & Related papers  (2025-12-30T08:16:20Z)
• ReLaX: Reasoning with Latent Exploration for Large Reasoning Models [11.506415241741601]
  We argue that latent dynamics underlying token generation encode a far richer computational structure for steering policy optimization.<n>We propose Reasoning with Latent eXploration (ReLaX), a paradigm that explicitly incorporates latent dynamics to regulate exploration and exploitation.
  arXiv  Detail & Related papers  (2025-12-08T13:48:33Z)
• Does More Inference-Time Compute Really Help Robustness? [50.47666612618054]
  We show that small-scale, open-source models can benefit from inference-time scaling.<n>We identify an important security risk, intuitively motivated and empirically verified as an inverse scaling law.<n>We urge practitioners to carefully weigh these subtle trade-offs before applying inference-time scaling in security-sensitive, real-world applications.
  arXiv  Detail & Related papers  (2025-07-21T18:08:38Z)
• Is Reasoning All You Need? Probing Bias in the Age of Reasoning Language Models [0.0]
  Reasoning Language Models (RLMs) have gained traction for their ability to perform complex, multi-step reasoning tasks.<n>While these capabilities promise improved reliability, their impact on robustness to social biases remains unclear.<n>We leverage the CLEAR-Bias benchmark to investigate the adversarial robustness of RLMs to bias elicitation.
  arXiv  Detail & Related papers  (2025-07-03T17:01:53Z)
• Lost at the Beginning of Reasoning [82.18834329384514]
  We show that the first reasoning step exerts a disproportionately large influence on the final prediction.<n>We propose an efficient sampling strategy that leverages a reward model to identify and retain high-quality first reasoning steps.<n>We introduce a new benchmark specifically constructed with deliberately flawed first reasoning steps to systematically evaluate model self-correction capabilities.
  arXiv  Detail & Related papers  (2025-06-27T09:53:57Z)
• Pitfalls of Rule- and Model-based Verifiers -- A Case Study on Mathematical Reasoning [26.717777746219635]
  We take mathematical reasoning as a case study and conduct a comprehensive analysis of various verifiers in both static evaluation and RL training scenarios.<n>First, we find that current open-source rule-based verifiers often fail to recognize equivalent answers presented in different formats across commonly used mathematical datasets, resulting in non-negligible false negative rates.<n>We investigate model-based verifiers as a potential solution to address these limitations.<n>While the static evaluation shows that model-based verifiers achieve significantly higher verification accuracy, further analysis and RL training results imply that they are highly susceptible to hacking, where they misclassify certain patterns
  arXiv  Detail & Related papers  (2025-05-28T10:28:41Z)
• Towards Optimal Adversarial Robust Reinforcement Learning with Infinity Measurement Error [9.473089575932375]
  Recent research highlights the challenges of achieving state-adversarial robustness.<n>We introduce the Intrinsic State-adversarial Markov Decision Process (ISA-MDP)<n>We show that improving DRL robustness does not necessarily compromise performance in natural environments.
  arXiv  Detail & Related papers  (2025-02-23T22:16:01Z)
• SAMBO-RL: Shifts-aware Model-based Offline Reinforcement Learning [9.88109749688605]
  Model-based offline reinforcement learning trains policies using pre-collected datasets and learned environment models.<n>This paper offers a comprehensive analysis that disentangles the problem into two fundamental components: model bias and policy shift.<n>We introduce Shifts-aware Model-based Offline Reinforcement Learning (SAMBO-RL), a practical framework that efficiently trains classifiers to approximate SAR for policy optimization.
  arXiv  Detail & Related papers  (2024-08-23T04:25:09Z)
• Time-Constrained Robust MDPs [28.641743425443]
  We introduce a new time-constrained robust MDP (TC-RMDP) formulation that considers multifactorial, correlated, and time-dependent disturbances. This study revisits the prevailing assumptions in robust RL and opens new avenues for developing more practical and realistic RL applications.
  arXiv  Detail & Related papers  (2024-06-12T16:45:09Z)
• Latent State Marginalization as a Low-cost Approach for Improving Exploration [79.12247903178934]
  We propose the adoption of latent variable policies within the MaxEnt framework. We show that latent variable policies naturally emerges under the use of world models with a latent belief state. We experimentally validate our method on continuous control tasks, showing that effective marginalization can lead to better exploration and more robust training.
  arXiv  Detail & Related papers  (2022-10-03T15:09:12Z)
• False Correlation Reduction for Offline Reinforcement Learning [115.11954432080749]
  We propose falSe COrrelation REduction (SCORE) for offline RL, a practically effective and theoretically provable algorithm. We empirically show that SCORE achieves the SoTA performance with 3.1x acceleration on various tasks in a standard benchmark (D4RL)
  arXiv  Detail & Related papers  (2021-10-24T15:34:03Z)
• Policy Smoothing for Provably Robust Reinforcement Learning [109.90239627115336]
  We study the provable robustness of reinforcement learning against norm-bounded adversarial perturbations of the inputs. We generate certificates that guarantee that the total reward obtained by the smoothed policy will not fall below a certain threshold under a norm-bounded adversarial of perturbation the input.
  arXiv  Detail & Related papers  (2021-06-21T21:42:08Z)
• Provably Good Batch Reinforcement Learning Without Great Exploration [51.51462608429621]
  Batch reinforcement learning (RL) is important to apply RL algorithms to many high stakes tasks. Recent algorithms have shown promise but can still be overly optimistic in their expected outcomes. We show that a small modification to Bellman optimality and evaluation back-up to take a more conservative update can have much stronger guarantees.
  arXiv  Detail & Related papers  (2020-07-16T09:25:54Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.