CRANE: Knowledge Editing for Reasoning MLLMs

• URL: http://arxiv.org/abs/2606.09033v1
• Date: Mon, 08 Jun 2026 05:01:11 GMT
• Title: CRANE: Knowledge Editing for Reasoning MLLMs
• Abstract summary: multimodal large language models (MLLMs) generate explicit chain-of-thought (CoT) reasoning before producing answers.<n>On reasoning MLLMs, methods that generalize (FT, LoRA) trigger format collapse, while methods without deep modification cannot generalize.<n>We propose CRANE, a retrieval-augmented framework that requires no per-edit parameter modification.
• Score: 37.084419383244814
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: The emergence of reasoning multimodal large language models (MLLMs), which generate explicit chain-of-thought (CoT) reasoning before producing answers, has introduced a new challenge for knowledge editing: methods that appear successful under traditional metrics (teacher-forcing accuracy up to 100%) can fail severely when the model's reasoning process is examined (Grounded Success as low as 0%). We identify three failure modes: (1) Structural Collapse, where weight-modifying methods destroy the CoT format; (2) Cognitive Dissonance, where the model's reasoning chain actively rejects the injected edit fact based on visual evidence; and (3) Shallow Internalization, where methods succeed on exact queries but fail on rephrase or multi-hop variants. On reasoning MLLMs, these modes interact: methods that generalize (FT, LoRA) trigger format collapse, while methods without deep modification cannot generalize. To expose these failures, we propose a CoT-aware evaluation protocol and construct ReasonEdit-Bench, with conflict stratification, multi-level probes, and multi-hop portability tests. We propose CRANE, a retrieval-augmented framework that requires no per-edit parameter modification. CRANE combines a modality-aware dual-library retrieval system with a two-phase training strategy: Supervised Fine-Tuning (SFT) for structural initialization, followed by GRPO with a Cognitive Routing Reward that trains the model to arbitrate between visual priors and injected edit facts. On ReasonEdit-Bench, CRANE achieves 96.9% Grounded Success on conflict scenarios and 96.9% intermediate entity usage in multi-hop chains, with 97.6% text-locality and 68.1% image-locality Edit Independence. On the out-of-distribution MMEVOKE benchmark, CRANE reaches 87.0% under gold retrieval.

Related papers

• AnE: Pushing the Reasoning Frontier of Multimodal LLMs via Anchor Evolution [61.593935260052795]
  Post-training viaSupervised Fine-Tuning (SFT) and Reinforcement Learning (RL) is crucial for enhancing reasoning in Multimodal Large Language Models (MLLMs)<n>Existing paradigms often reach a performance bottleneck due to the limitations of static data.<n>We propose Anchor Evolution (AnE), a new paradigm that integrates truth-anchored data curation and model evolution.
  arXiv  Detail & Related papers  (2026-05-25T08:26:34Z)
• ReLoop: Structured Modeling and Behavioral Verification for Reliable LLM-Based Optimization [6.572539312871392]
  Large language models (LLMs) can translate natural language into optimization code, but silent failures pose a critical risk.<n>We introduce ReLoop, addressing silent failures from two complementary directions.
  arXiv  Detail & Related papers  (2026-02-17T20:20:33Z)
• CRAFT: Calibrated Reasoning with Answer-Faithful Traces via Reinforcement Learning for Multi-Hop Question Answering [19.391824811629125]
  Retrieval-augmented generation (RAG) is widely used to ground Large Language Models (LLMs) for multi-hop question answering.<n> Reasoning in multi-hop QA is inherently complex due to multi-hop composition and is further destabilized by noisy retrieval.<n>We propose CRAFT, a reinforcement learning framework that trains models to perform faithful reasoning during response generation.
  arXiv  Detail & Related papers  (2026-02-01T17:33:39Z)
• VERGE: Formal Refinement and Guidance Engine for Verifiable LLM Reasoning [4.3414302048068745]
  We present a neurosymbolic framework that combines Large Language Models with SMT solvers to produce verification-guided answers.<n>We introduce three key innovations: (1) multi-model consensus via formal semantic equivalence checking, (2) semantic routing that directs different claim types to appropriate verification strategies, and (3) precise logical error localization via Minimal Correction Subsets.<n>With the GPT-OSS-120B model, VERGE demonstrates an average performance uplift of 18.7% at convergence across a set of reasoning benchmarks compared to single-pass approaches.
  arXiv  Detail & Related papers  (2026-01-27T20:59:11Z)
• CoG: Controllable Graph Reasoning via Relational Blueprints and Failure-Aware Refinement over Knowledge Graphs [53.199517625701475]
  CoG is a training-free framework inspired by Dual-Process Theory that mimics the interplay between intuition and deliberation.<n>CoG significantly outperforms state-of-the-art approaches in both accuracy and efficiency.
  arXiv  Detail & Related papers  (2026-01-16T07:27:40Z)
• KBQA-R1: Reinforcing Large Language Models for Knowledge Base Question Answering [64.62317305868264]
  We present textbfKBQA-R1, a framework that shifts the paradigm from text imitation to interaction optimization via Reinforcement Learning.<n>Treating KBQA as a multi-turn decision process, our model learns to navigate the knowledge base using a list of actions.<n>Experiments on WebQSP, GrailQA, and GraphQuestions demonstrate that KBQA-R1 achieves state-of-the-art performance.
  arXiv  Detail & Related papers  (2025-12-10T17:45:42Z)
• Erase to Improve: Erasable Reinforcement Learning for Search-Augmented LLMs [18.37387666170851]
  We propose Erasable Reinforcement Learning (ERL), a novel framework that transforms fragile reasoning into a robust process.<n>ERL explicitly identifies faulty steps, erases them, and regenerates reasoning in place, preventing defective logic from propagating through the reasoning chain.<n>Models trained with ERL, termed ESearch, achieve substantial improvements on HotpotQA, MuSiQue, 2Wiki, and Bamboogle.
  arXiv  Detail & Related papers  (2025-10-01T13:10:36Z)
• Keys to Robust Edits: from Theoretical Insights to Practical Advances [20.10464264597003]
  Large language models (LLMs) struggle with maintaining accurate knowledge due to conflicting/outdated parametric memories.<n>Our solution introduces textitRobust Edit Pathway (REP), a plug-and-play module that disentangles editing keys from native model representations.
  arXiv  Detail & Related papers  (2024-10-12T02:54:12Z)
• Disperse-Then-Merge: Pushing the Limits of Instruction Tuning via Alignment Tax Reduction [75.25114727856861]
  Large language models (LLMs) tend to suffer from deterioration at the latter stage ofSupervised fine-tuning process. We introduce a simple disperse-then-merge framework to address the issue. Our framework outperforms various sophisticated methods such as data curation and training regularization on a series of standard knowledge and reasoning benchmarks.
  arXiv  Detail & Related papers  (2024-05-22T08:18:19Z)
• Chain of Evidences and Evidence to Generate: Prompting for Context Grounded and Retrieval Augmented Reasoning [3.117335706912261]
  Chain of Evidences (CoE) and Evidence to Generate (E2G) are built upon two unique strategies.<n>Instead of unverified reasoning claims, our innovative approaches leverage the power of "evidence for decision making"<n>Our framework consistently achieves remarkable results across various knowledge-intensive reasoning and generation tasks.
  arXiv  Detail & Related papers  (2024-01-11T09:49:15Z)
• DCR: Divide-and-Conquer Reasoning for Multi-choice Question Answering with LLMs [9.561022942046279]
  We propose Divide and Conquer Reasoning (DCR) to enhance the reasoning capability of large language models (LLMs) We first categorize questions into two subsets based on confidence score ($mathcalCS$), which is estimated by statistical frequency of generated answers. In particular, we first categorize questions into two subsets based on confidence score ($mathcalCS$), which is estimated by statistical frequency of generated answers.
  arXiv  Detail & Related papers  (2024-01-10T14:38:46Z)
• RCOT: Detecting and Rectifying Factual Inconsistency in Reasoning by Reversing Chain-of-Thought [56.558892336235914]
  Reversing Chain-of-Thought (RCoT) is a novel method to improve large language models' reasoning abilities. RCoT automatically detects and rectifys factual inconsistency in generated solutions. We show that manually written fine-grained feedback can dramatically improve LLMs' reasoning abilities.
  arXiv  Detail & Related papers  (2023-05-19T08:02:52Z)

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.