Research map / Vision-language and multimodal / Multimodal LLMs

MIND: Multi-rationale INtegrated Discriminative Reasoning Framework for Multi-modal Large Models

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs · 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 MIND, a multi-rationale discriminative training framework (RAD data expansion, two-phase P2CL curriculum, and margin-based MCA contrastive loss) that fine-tunes a T5-based multimodal CoT model and reports SOTA on ScienceQA, A-OKVQA, and M3CoT.

Problem

Single-rationale supervision in multimodal chain-of-thought training leaves models rigid, logically fragile, and easily misled by flawed explanations, with limited ability to discriminate and self-correct reasoning.

Approach

Expand each training sample with many LLM-generated positive rationales (meaning-preserving paraphrases) and negative rationales (meaning-reversed edits); train first on positives, then condition on positive or negative rationales to jointly generate the answer and a corrected positive rationale, and add a margin contrastive loss that pulls hard positives together and pushes hard negatives apart.

Key results

• MIND_base reaches 92.29% average accuracy on ScienceQA, reported as SOTA and above Multimodal-CoT_base (85.31%) by 6.98 points.
• MIND_base achieves 70.6% on A-OKVQA multiple-choice vs 50.6% for Multimodal-CoT_base.
• On M3CoT, MIND_large reaches 61.56% total, surpassing GPT-4V (56.95%) with a 738M model; ablation shows P2CL alone adds 1.86 points on ScienceQA.

Surprising findings

• MCA loss alone yields only a 0.07-point gain on ScienceQA, while combining it with P2CL gives a larger jump, which the authors describe as synergistic.
• Using rationales from a stronger generator (Deepseek-R1) did not outperform the Qwen3-235B-A22B source, and a mix of sources performed best.

Limitations

• Requires offline construction of large augmented rationale datasets (up to 21M rationales), adding data preparation cost.
• Validation limited to VQA-style benchmarks.
• Still prone to perceptual bias and reasoning errors on fine-grained visual understanding; quality bounded by original rationales.

Experimental setup

Datasets: ScienceQA A-OKVQA M3CoT

Benchmarks: ScienceQA A-OKVQA M3CoT

Metrics: accuracy rationale quality score (correctness, relevance, coherence,

Baselines: Multimodal-CoT MC-CoT DPMM-CoT Multimodal-T-SciQ Chameleon GPT-4V LLaVA GPT-4 w/ CoT

Builds on: Multimodal-CoT MC-CoT UnifiedQA BLIP-2

Closest papers by meaning

• MDK12-Bench: A Multi-Discipline Benchmark for Evaluating Reasoning in Multimodal Large Language Models (AAAI 2026 · similarity 0.87)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.87)
• Rationale-Enhanced Decoding for Multi-modal Chain-of-Thought (CVPR 2026 · similarity 0.86)
• LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.86)
• Revisual-R1: Advancing Multimodal Reasoning From Optimized Cold Start to Staged Reinforcement Learning (ICLR 2026 · similarity 0.86)
• Toward Cognitive Supersensing in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.86)