Research map / Vision-language and multimodal / Multimodal LLMs

Perception Without Engagement: Dissecting the Causal Discovery Deficit in LMMs

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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.

Introduces ProCauEval, a perturbation-based protocol with five configurations to diagnose textual-prior reliance in LMM causal discovery over videos, evaluates 17 LMMs, and proposes ADPO, a GRPO extension that repels the policy from a visually-corrupted prior-only teacher.

Problem

LMMs answer video causal discovery questions largely via textual priors rather than visual evidence, and existing benchmarks report only outcome accuracy, not the sources of this failure.

Approach

Perturbation-based diagnosis (fake events, fake video, blank video, captioning) decomposes modality contributions, and ADPO adds a KL repulsion term between policy distributions conditioned on original versus visually corrupted video (stop-gradient on the corrupted branch) on top of GRPO.

Key results

• Across 17 LMMs, mean R-Avg falls from 33.9% under Veridicality to 2.4% under Plausibility (about 88.7% relative drop).
• ADPO with frame shuffling on Qwen3-VL-8B raises Veridicality R-Avg from 36.4 (base) / 41.3 (GRPO) to 43.1 and lowers Plausibility FR from 52.1 (GRPO) to 42.2.
• ADPO improves Video-MME, LongVideoBench, MMVU and MVBench over Qwen3-VL-8B, GRPO and DAPO (e.g., LongVideoBench 62.3 vs 60.5 for GRPO).

Surprising findings

• Thinking/reasoning-oriented post-training amplifies textual prior reliance, and several thinking models score higher R-Avg under Incongruence than under Veridicality.
• Higher Veridicality accuracy correlates with greater fragility under perturbation (Pearson r = 0.832 between R-Avg and FR across 17 LMMs).

Experimental setup

Datasets: MECD Video-R1-260k

Benchmarks: ProCauEval Video-MME LongVideoBench MMVU MVBench

Metrics: R-Avg Acc Flip Rate Non-Flip Rate GPT Score

Baselines: GRPO DAPO LLaVA-OneVision Qwen3-VL InternVL-3.5 MiMo-VL GPT-5.1 Video-R1 VideoChat-R1.5 Keye-VL-1.5 LongVILA-R1 GLM-4.1V-Thinking

Builds on: GRPO DAPO MECD Video-R1

Closest papers by meaning

• On the Faithfulness of Visual Thinking: Measurement and Enhancement (NeurIPS 2026 · similarity 0.88)
• When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.88)
• From Prompts to Tokens: Internalizing Causal Supervision in Vision-Language Model for Multi-Image Causal Reasoning (ICML 2026 · similarity 0.87)
• Beyond Perceptual Shortcuts: Causal-Inspired Debiasing Optimization for Generalizable Video Reasoning in Lightweight MLLMs (CVPR 2026 · similarity 0.87)
• SciVideoBench: Benchmarking Scientific Video Reasoning in Large Multimodal Models (ICML 2026 · similarity 0.87)
• Spotlight on Token Perception for Multimodal Reinforcement Learning (ICLR 2026 · similarity 0.87)