A Rolling Stone Gathers No Moss: Adaptive Policy Optimization for Stable Self-Evaluation in Large Multimodal Models

• URL: http://arxiv.org/abs/2508.09155v1
• Date: Tue, 05 Aug 2025 07:54:01 GMT
• Title: A Rolling Stone Gathers No Moss: Adaptive Policy Optimization for Stable Self-Evaluation in Large Multimodal Models
• Authors: Wenkai Wang, Hongcan Guo, Zheqi Lv, Shengyu Zhang,
• Abstract summary: We propose AdaPO, an online reinforcement learning framework capable of adaptively adjusting training objective in real time.<n>We show that our method significantly enhances both direct reasoning and self-evaluation capability.
• Score: 4.417707977122247
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Self-evaluation, a model's ability to assess the correctness of its own output, is crucial for Large Multimodal Models (LMMs) to achieve self-improvement in multi-turn conversations, yet largely absent in foundation models. Recent work has employed reinforcement learning (RL) to enhance self-evaluation; however, its fixed reward mechanism suffers from reward hacking when optimizing multiple training objectives, leading to model collapse. In this paper we propose AdaPO, an online reinforcement learning framework capable of adaptively adjusting training objective in real time according to the current training state for each task. Specifically, to mitigate reward hacking , AdaPO introduces an Adaptive Reward Model (ARM) and a Reward Aware Dynamic KL Regularization mechanism. ARM assesses the task's training state from the distribution of model generated multi-turn trajectories' performance. Reward Aware Dynamic KL replaces a fixed penalty with dynamic coefficients which is modulated by the reward gap between different multi-turn situations. Notably, our method automatically and smoothly adjusts its learning focus based on sub-tasks' training progress without manual intervention. Extensive experiments over 8 benchmarks and various models show that our method significantly enhances both direct reasoning and self-evaluation capability. We will release our code to contribute to the community.

Related papers

• Reinforcement World Model Learning for LLM-based Agents [60.65003139516272]
  Reinforcement World Model Learning (RWML) is a self-conditioned method that learns action-supervised world models for LLM-based agents.<n>Our method aligns simulated next states produced by the model with realized next states observed from the environment.<n>We evaluate our method on ALFWorld and $2$ Bench and observe significant gains over the base model, despite being entirely self-supervised.
  arXiv  Detail & Related papers  (2026-02-05T16:30:08Z)
• OpenReward: Learning to Reward Long-form Agentic Tasks via Reinforcement Learning [41.49024599460379]
  Reward models (RMs) have become essential for aligning large language models (LLMs)<n>We introduce OpenRM, a tool-augmented long-form reward model that judges open-ended responses by invoking external tools to gather relevant evidence.<n>Experiments on three newly-collected datasets and two widely-used benchmarks demonstrate that OpenRM substantially outperforms existing reward modeling approaches.
  arXiv  Detail & Related papers  (2025-10-28T17:02:46Z)
• Confidence as a Reward: Transforming LLMs into Reward Models [54.98336080630691]
  Confidence-as-a-Reward (CRew) is a training-free method that utilizes token-level confidence in the model's final answers as a proxy for reward.<n>We show that CRew outperforms existing training-free reward approaches on the MATH500 and RewardMATH benchmarks.<n>We propose CRew-DPO, a training strategy that constructs preference data from confidence scores combined with correctness signals.
  arXiv  Detail & Related papers  (2025-10-15T12:51:47Z)
• Beyond Monolithic Rewards: A Hybrid and Multi-Aspect Reward Optimization for MLLM Alignment [1.8552770604791606]
  We propose a hybrid reward modeling framework that integrates complementary reward paradigms.<n>We show consistent improvements across different multimodal benchmarks when applying hybrid and multi-aspect reward modeling.<n>Our best performing model in the 3B family achieves an overall average improvement of 9.5% across general and math reasoning tasks.
  arXiv  Detail & Related papers  (2025-10-06T18:53:23Z)
• Activation Reward Models for Few-Shot Model Alignment [77.37511364793515]
  We introduce Activation Reward Models (Activation RMs)<n>Activation RMs leverage activation steering to construct well-aligned reward signals using minimal supervision and no additional model finetuning.<n>We demonstrate the effectiveness of Activation RMs in mitigating reward hacking behaviors, highlighting their utility for safety-critical applications.
  arXiv  Detail & Related papers  (2025-07-02T05:10:29Z)
• Toward Evaluative Thinking: Meta Policy Optimization with Evolving Reward Models [20.30272221358623]
  We introduce a framework that integrates a meta-reward model that dynamically refines the reward model's prompt throughout training.<n>In MPO, the meta-reward model monitors the evolving training context and continuously adjusts the reward model's prompt to maintain high alignment.<n>It yields performance on par with or better than models guided by extensively hand-crafted reward prompts.
  arXiv  Detail & Related papers  (2025-04-28T18:02:35Z)
• Reinforcement Learning for Machine Learning Model Deployment: Evaluating Multi-Armed Bandits in ML Ops Environments [0.0]
  We investigate whether reinforcement learning (RL)-based model management can manage deployment decisions more effectively.<n>Our approach enables more adaptive production environments by continuously evaluating deployed models and rolling back underperforming ones in real-time.<n>Our findings suggest that RL-based model management can improve automation, reduce reliance on manual interventions, and mitigate risks associated with post-deployment model failures.
  arXiv  Detail & Related papers  (2025-03-28T16:42:21Z)
• Multiagent Finetuning: Self Improvement with Diverse Reasoning Chains [114.76612918465948]
  Large language models (LLMs) have achieved remarkable performance in recent years but are fundamentally limited by the underlying training data.<n>We propose a complementary approach towards self-improvement where finetuning is applied to a multiagent society of language models.
  arXiv  Detail & Related papers  (2025-01-10T04:35:46Z)
• Diving into Self-Evolving Training for Multimodal Reasoning [36.70979791148913]
  Self-evolving trainin has emerged as a key approach for complex reasoning tasks.<n>This paper reframes self-evolving training for multimodal reasoning through the lens of reinforcement learning.<n>We propose M-STAR, a framework that achieves consistent performance gains across models of varying sizes and diverse benchmarks.
  arXiv  Detail & Related papers  (2024-12-23T10:18:41Z)
• Training Language Models to Self-Correct via Reinforcement Learning [98.35197671595343]
  Self-correction has been found to be largely ineffective in modern large language models (LLMs) We develop a multi-turn online reinforcement learning approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. We find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on MATH and HumanEval.
  arXiv  Detail & Related papers  (2024-09-19T17:16:21Z)
• Jointly Training and Pruning CNNs via Learnable Agent Guidance and Alignment [69.33930972652594]
  We propose a novel structural pruning approach to jointly learn the weights and structurally prune architectures of CNN models. The core element of our method is a Reinforcement Learning (RL) agent whose actions determine the pruning ratios of the CNN model's layers. We conduct the joint training and pruning by iteratively training the model's weights and the agent's policy.
  arXiv  Detail & Related papers  (2024-03-28T15:22:29Z)
• Model-Based Reinforcement Learning with Multi-Task Offline Pretraining [59.82457030180094]
  We present a model-based RL method that learns to transfer potentially useful dynamics and action demonstrations from offline data to a novel task. The main idea is to use the world models not only as simulators for behavior learning but also as tools to measure the task relevance. We demonstrate the advantages of our approach compared with the state-of-the-art methods in Meta-World and DeepMind Control Suite.
  arXiv  Detail & Related papers  (2023-06-06T02:24:41Z)
• Distributional Reward Estimation for Effective Multi-Agent Deep Reinforcement Learning [19.788336796981685]
  We propose a novel Distributional Reward Estimation framework for effective Multi-Agent Reinforcement Learning (DRE-MARL) Our main idea is to design the multi-action-branch reward estimation and policy-weighted reward aggregation for stabilized training. The superiority of the DRE-MARL is demonstrated using benchmark multi-agent scenarios, compared with the SOTA baselines in terms of both effectiveness and robustness.
  arXiv  Detail & Related papers  (2022-10-14T08:31:45Z)

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.