AnE: Pushing the Reasoning Frontier of Multimodal LLMs via Anchor Evolution

2026-05-25

AI Summary

Existing self-evolution approaches for multimodal large language models are highly susceptible to low-quality synthetic data, often leading to cognitive drift and hallucinatory reasoning. To address this, this work proposes the Anchor-based Evolution (AnE) paradigm, which leverages ground-truth anchors to filter high-fidelity training samples, employs trajectory backtracking to identify model failure boundaries, and utilizes a scaffolding-stripping strategy to internalize robust reasoning capabilities. By integrating supervised fine-tuning, reinforcement learning, and retrieval from a curated ground-truth database, AnE ensures reasoning faithfulness while substantially enhancing performance. The method achieves state-of-the-art results across eight mainstream multimodal reasoning benchmarks, yielding an average improvement of 10.3%.

Abstract

Post-training via Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) is crucial for enhancing reasoning in Multimodal Large Language Models (MLLMs), yet existing paradigms often reach a performance bottleneck due to the limitations of static data. While current methods leverage self-reflection or self-evolution to push these boundaries, they still suffer from cognitive drift and hallucinated reasoning paths caused by low-quality synthetic data. To address these challenges, we propose Anchor Evolution (AnE), a new paradigm that integrates truth-anchored data curation and model evolution, achieving faithful and steady performance gains at the reasoning frontier. Specifically, we propose Truth Anchor Expansion, which pinpoints the model failing frontier via trajectory rollouts and leverages ground-truth databases to retrieve high-fidelity anchors for faithful data curation. Subsequently, we introduce the Scaffold-Stripping Mechanism to internalize reasoning capabilities. This mechanism first anchors reasoning paths via scaffold-augmented supervision to mitigate the learning complexity and distribution drift of direct SFT on raw data, then leverages RL to strip the scaffold template, thereby effectively transitioning the reasoning paths into intrinsic model capabilities. Experimental results on multimodal reasoning benchmarks show that our method substantially advances the model performance frontier, improving the base model by 10.3\% across eight multimodal benchmarks and achieving state-of-the-art results. The code will be made publicly available.

Problem

• Multimodal LLMs
• reasoning
• cognitive drift
• hallucinated reasoning
• static data

Innovation

• Anchor Evolution
• Truth Anchor Expansion
• Scaffold-Stripping Mechanism
• Multimodal Reasoning
• Model Evolution