Research map / Vision-language and multimodal / Multimodal LLMs

Boosting Reasoning in Large Multimodal Models via Activation Replay

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Activation Replay, a training-free test-time method that learns zero-initialized visual-token offsets to make low-entropy input activations of RLVR-tuned LMMs match their base counterparts, improving multimodal reasoning and Pass@K.

Problem

Post-training with RLVR changes LMM internals in poorly understood ways, and it is unclear how these input-activation shifts relate to multimodal reasoning and whether they can be exploited without expensive policy optimization.

Approach

Using logit-lens analysis, RLVR is found to shift low-entropy input activations more than high-entropy ones; Activation Replay learns test-time additive offsets on visual tokens, optimized with a KL term restricted to low-entropy base-model positions, so RLVR activations mimic the base model's low-entropy distributions.

Key results

• Activation Replay improves MM-Eureka-Qwen-7B on MathVerse (Vision Only) from 45.1 to 47.7 and on LogicVista from 49.2 to 51.0 (Table 2).
• On DeepEyes-7B, replay raises VisualProbe-Medium from 29.8 to 32.8 and HRBench-4K from 53.7 to 56.3.
• On Video-R1-7B, replay raises VideoMMMU from 49.8 to 53.8 and Video-Holmes from 36.5 to 40.9; Pass@16 on MathVision mini for VLAA-Thinker rises from 84.5 to 89.5.

Surprising findings

• Replaying high-entropy activations from the base model instead of low-entropy ones degrades performance, as shown in the intervention study.
• Direct cross-model intervention with base activations is less effective than learnable token manipulation because base and RLVR representations are not aligned.

Limitations

• Gains are modest on several benchmarks (often 0.5 to 2 points), and some settings show drops (e.g., DynaMath for MM-Eureka-32B).
• Mechanism linking low-entropy activation shifts to reasoning is supported by correlational perturbation studies and the authors leave the underlying cause open.
• Test-time optimization adds inference cost that varies by task, and the hyperparameters tau and alpha are tuned per setting.

Experimental setup

Datasets: MathVerse MathVision MathVista DynaMath WeMath LogicVista MMMU MMMUPro HR-Bench VisualProbe VideoMMMU MLVU

Benchmarks: MathVerse (Vision Only) MathVision MathVista DynaMath WeMath LogicVista MMMU MMMUPro HR-Bench VisualProbe VideoMMMU MLVU

Metrics: accuracy Pass@K relative time cost

Baselines: MMR1-Math-v0-7B MM-Eureka-Qwen-7B MM-Eureka-Qwen-32B VL-Rethinker-7B DeepEyes-7B Video-R1-7B VLAA-Thinker Qwen2.5-VL-7B

Builds on: GRPO logit lens DeepEyes Video-R1 MM-Eureka

Closest papers by meaning

• CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal Reasoning (CVPR 2026 · similarity 0.90)
• R-C2: Cycle-Consistent Reinforcement Learning Improves Multimodal Reasoning (CVPR 2026 · similarity 0.90)
• SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
• Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models (CVPR 2026 · similarity 0.89)
• Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.89)
• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.89)