💻 computer science

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

The paper proposes Anchor Evolution (AnE), a novel paradigm that combines Truth Anchor Expansion for high-fidelity data curation and a Scaffold-Stripping Mechanism to internalize reasoning capabilities, thereby overcoming cognitive drift and achieving state-of-the-art performance on multimodal reasoning benchmarks.

Original authors: Zehao Wang, Yihan Zeng, Zidong Gong, Yuanfan Guo, Feng Zhu, Hongzhi Zhang, Wei Zhang, Wangmeng Zuo

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zehao Wang, Yihan Zeng, Zidong Gong, Yuanfan Guo, Feng Zhu, Hongzhi Zhang, Wei Zhang, Wangmeng Zuo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Teaching a Robot to Think Better

Imagine you are trying to teach a very smart robot (a Multimodal Large Language Model) to solve complex puzzles, like math problems with diagrams or logic riddles. You want it to get better at reasoning, not just memorizing answers.

The paper argues that the current ways of teaching these robots have two big problems:

The Static Library: If you just give the robot a fixed set of textbooks (static data), it hits a ceiling. Once it learns everything in those books, it can't get any smarter because the books don't change to match its growing skills.
The Hallucinating Tutor: If you let the robot teach itself by making up new practice problems (self-evolution), it often starts lying to itself. It creates fake logic and wrong answers that sound convincing but are actually nonsense. This is called "cognitive drift."

The Solution: The authors propose a new method called Anchor Evolution (AnE). Think of it as a "Truth-Anchored" training camp that keeps the robot grounded in reality while pushing it to its limits.

How AnE Works: The Three-Step Training Camp

The method works in a loop, like a cycle of practice, review, and mastery.

1. Finding the "Failing Frontier" (The Weak Spot Detector)

First, the system asks the robot to try solving a bunch of problems. It doesn't just look at whether the robot got the answer right or wrong; it looks at how the robot thought.

The Analogy: Imagine a coach watching an athlete run a race. The coach doesn't just check the finish time. They notice exactly where the athlete stumbled or got confused.
The Paper's Claim: The system identifies the specific types of problems where the robot consistently fails. These are the "Failing Frontier" areas—the edge of what the robot currently knows.

2. "Truth Anchor Expansion" (The Reality Check)

Once the robot's weak spots are found, the system needs to create new practice problems to fix them.

The Problem with Old Methods: Previous methods would ask a "Teacher AI" to make up new problems. But the Teacher AI might make mistakes or invent fake facts, leading the student robot down a rabbit hole of lies.
The AnE Solution: Instead of making things up, AnE goes to a massive, verified library of real-world data (a "Ground-Truth Database"). It searches for real, correct examples that match the specific type of mistake the robot made.
The Analogy: If a student keeps failing at "adding fractions," a bad teacher might invent a fake rule. A good teacher (AnE) goes to a library of real math textbooks, finds actual, verified fraction problems, and says, "Here are real examples of exactly what you struggled with." These real examples are the "Truth Anchors." They keep the training grounded in reality.

3. The "Scaffold-Stripping Mechanism" (Training Wheels On, Then Off)

Now the robot has real problems, but they might still be too hard for it to solve alone immediately.

Step A: Scaffolding (The Training Wheels): The system takes the "Teacher AI" and asks it to give a helpful hint or a step-by-step guide for the problem. The robot practices solving the problem with this help. This is called "Scaffold-Augmented Supervision."
- Analogy: It's like a driving instructor sitting in the passenger seat, saying, "Turn the wheel here, then press the gas." The student learns the logic of the move with support.
Step B: Stripping (Removing the Wheels): Once the robot has practiced with the hints, the system removes the hints. The robot is then tested using Reinforcement Learning (RL) to solve the same problems on its own.
- Analogy: The instructor gets out of the car. The student must now drive the same route alone. If they succeed, it proves they actually learned the skill, not just memorized the instructor's voice.

Why This is Better (The Results)

The paper tested this method on a robot called Qwen2.5-VL-7B.

The Result: By using this "Anchor Evolution" loop, the robot improved its reasoning skills by 10.3% across eight different difficult benchmarks (like MathVision and EMMA).
The Comparison:
- Static Training: The robot hit a wall and stopped improving.
- Self-Evolution: The robot got confused, started hallucinating (making up facts), and actually got worse over time.
- AnE: The robot kept getting smarter, round after round, without losing its mind.

Summary in One Sentence

Anchor Evolution is a training method that finds exactly where a robot is failing, pulls real, verified examples from a database to fix those specific weaknesses, and then uses a "training wheels" approach to teach the robot how to solve those problems on its own, ensuring it learns true reasoning skills without getting lost in fake logic.

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

Problem Statement

Post-training Multimodal Large Language Models (MLLMs) via Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) is essential for enhancing reasoning capabilities. However, existing paradigms face significant bottlenecks:

Static Post-Training: Reliance on fixed datasets limits the model's ability to adapt to its evolving weaknesses, leading to diminishing returns as the model's capability frontier shifts.
Self-Evolution: Methods that generate synthetic data to create closed-loop training cycles often suffer from cognitive drift and hallucinated reasoning paths. When the training signal is dominated by low-quality self-synthetic data, errors accumulate across iterations, degrading performance rather than improving it.

The core challenge is to achieve scalable model evolution that targets the model's specific failure points while maintaining high-fidelity, truthful supervision to prevent hallucination.

Methodology: Anchor Evolution (AnE)

The authors propose Anchor Evolution (AnE), an iterative post-training framework that integrates truth-anchored data curation with model evolution. The framework operates through three integrated stages repeated in an iterative cycle:

1. Failing-Frontier Discovery

The process begins by identifying the specific regions where the model underperforms.

Rollouts: The model performs $K$ stochastic rollouts on a seed dataset.
Data Routing: Samples are categorized based on success counts ( $c$ $c$ ):
- Redundant ( $c=K$ ): Discarded to save compute.
- Volatile ( $1 \le c < K$ ): Retained for stability.
- Failing Frontier ( $c=0$ ): Samples where the model consistently fails. These represent the critical gaps in the model's capabilities.
Diagnostic Analysis: A teacher model analyzes the failure trajectories of the "Failing Frontier" samples to extract:
- Search Keywords ( $k_x$ ): 3-5 nouns related to the failure, used to retrieve external data.
- Actionable Hints ( $h_x$ ): 1-3 sentences providing structured guidance to solve the problem.

2. Truth Anchor Expansion

To avoid the hallucination risks of synthetic data generation, AnE retrieves verified samples from an external, ground-truth database.

Retrieval: Using the search keywords ( $k_x$ ) derived from the failure analysis, the system retrieves high-fidelity "truth anchors" from a large-scale ground-truth database (e.g., FineVision).
Curation: These anchors serve as verified references, ensuring the training data is grounded in real-world knowledge rather than model-generated speculation. This step expands the dataset specifically targeting the identified logical deficiencies.

3. Scaffold-Stripping Mechanism

This stage bridges the gap between the model's current inability to solve a problem and its eventual mastery of it.

Scaffold-Augmented SFT: The actionable hints ( $h_x$ ) are appended to the input questions to guide the model toward correct reasoning trajectories. Successful guided trajectories are collected, and the hints are moved from the input to the output prefix, creating "scaffold-augmented" training data. This allows the model to learn correct reasoning paths with external support.
RL Stripping: The model is then trained via Reinforcement Learning on the same problems but without the hints. The goal is to "strip" the scaffold, forcing the model to internalize the reasoning logic and generate correct answers independently.

Key Contributions

Anchor Evolution Framework: A novel post-training paradigm that tightly couples truth-anchored data curation with model evolution, achieving steady gains at the reasoning frontier.
Truth Anchor Expansion: A mechanism that pinpoints model failures via trajectory rollouts and retrieves verified anchors from ground-truth databases. This prevents the cognitive drift and hallucinations associated with purely synthetic self-evolution.
Scaffold-Stripping Mechanism: A two-step process using scaffold-augmented supervision (SFT) followed by RL to remove the scaffolds. This enables the model to internalize external guidance as an intrinsic capability rather than merely imitating it.

Experimental Results

The authors evaluated AnE using Qwen2.5-VL-7B as the base model across eight multimodal reasoning benchmarks (including MathVista, MathVision, EMMA, and MMMU).

Performance Gains: AnE improved the base model by an average of 10.3% across the eight benchmarks.
State-of-the-Art: The method achieved state-of-the-art results on challenging datasets like MathVision and EMMA.
Iterative Improvement: Unlike static methods that plateau or self-evolution methods that degrade in later rounds due to hallucination, AnE demonstrated continuous improvement over three evolution rounds.
Hallucination Reduction: Compared to self-evolution baselines, AnE produced significantly fewer hallucinated samples (0.2–0.8% vs. 6.5–9.3%), confirming the effectiveness of truth-anchored curation.
Ablation Studies: Experiments confirmed that both Truth Anchor Expansion and the Scaffold-Stripping Mechanism contribute independently, but their combination yields the best performance. The method also showed robustness to database scale and generalization to other model families (e.g., InternVL3.5).

Significance and Claims

The paper claims that Anchor Evolution addresses the fundamental limitations of current post-training strategies. By grounding data evolution in verified real-world knowledge (Truth Anchors) and systematically internalizing reasoning capabilities (Scaffold Stripping), AnE enables faithful and steady performance gains.

The authors argue that this approach overcomes the "performance bottleneck" of static data and the "cognitive drift" of synthetic self-evolution. The significance lies in providing a scalable, reliable pathway for pushing the reasoning frontier of MLLMs, ensuring that improvements are driven by genuine capability expansion rather than superficial imitation or hallucinated reasoning paths. The work suggests that effective model evolution requires both adaptability to the model's failing frontier and reliable, verified knowledge sources.