Mitigating Evidence Suppression: Bi-level Active Evidence Injection for Educational Video Understanding
sigir26-062 · Full research paper · Cheng Liu, Yiping Wang, Quanlong Guan, Chaobo He, Xingyu Zhu, Liangda Fang
Video and Temporal Understanding · 50% Multimodal Evaluation and Benchmarks · 50%
Abstract
Large Vision--Language Models (LVLMs) frequently fail on knowledge-intensive educational video QA despite the presence of requisite visual evidence. Through region-level analysis, we identify a systematic evidence-suppression pattern: task-critical tokens (e.g., diagrams) exhibit lower representational energy than distractors at the encoder output, rendering them prone to persistent attention neglect during decoding. While a controlled study shows that performance is sensitive to the strength and amount of injected candidate evidence signals, we find that rigid heuristics are insufficient due to sensitivity to token quality. To address this, we propose Bi-level Active Evidence Injection (BAEI), a decoding-time intervention that keeps the LVLM backbone frozen. BAEI employs a lightweight Injection Policy Network (IPN), optimized via GRPO, to dynamically select candidate evidence tokens and predict their token-wise injection strengths. The framework operates on two levels: increasing the contribution of candidate evidence-related signals in shallow layers and performing adaptive correction in deep layers triggered by predictive entropy. Experiments on educational benchmarks demonstrate consistent gains, validating decoding-time evidence intervention as an effective solution for factual alignment. The code is available at https://github.com/diojojolc-cell/BAEI.
Assigned reviewers
| # | Reviewer | Match | Why (closest evidence publications in the taxonomy) | |
|---|---|---|---|---|
| 1 | Zihao Yue | 82 | 100 OpenMMEgo: Enhancing Egocentric Understanding for LMMs with Open Weights and Data (2025) Video and Temporal Understanding 100 Movie101v2: Improved Movie Narration Benchmark (2025) Video and Temporal Understanding 100 Movie101: A New Movie Understanding Benchmark (2023) Video and Temporal Understanding |
|
| 2 | Haiyang Xu 0001 | 65 | 67 mPLUG-Owl3: Towards Long Image-Sequence Understanding in Multi-Modal Large Language Models (2025) Video and Temporal Understanding 67 TinyChart: Efficient Chart Understanding with Program-of-Thoughts Learning and Visual Token Merging (2024) Multimodal Evaluation and Benchmarks 67 Hallucination Augmented Contrastive Learning for Multimodal Large Language Model (2024) Multimodal Evaluation and Benchmarks |
|
| 3 | Nancy F. Chen | 45 | 71 Multimodal Dialogue State Tracking (2022) Conversational Search and Retrieval 67 Topic-Aware Pointer-Generator Networks for Summarizing Spoken Conversations (2019) Multimodal Evaluation and Benchmarks 67 Controllable Neural Dialogue Summarization with Personal Named Entity Planning (2021) Multimodal Evaluation and Benchmarks |
Area chair
| Area chair | Match | |
|---|---|---|
| Hao Fei 0001 | 52 |
Area-chair candidates (6)
| Xiaoye Qu | 53 | 15 papers | |
| Hao Fei 0001 assigned | 52 | 4 papers | |
| Yi R. Fung 0001 | 39 | 10 papers | |
| Kam-Fai Wong | 37 | 6 papers | |
| Xun Yang 0001 | 37 | 4 papers | |
| Hongjin Qian | 32 | 12 papers |
Candidates from the panel ranked by match score
| # | Reviewer | Match | Load | Why | |
|---|---|---|---|---|---|
| 1 | Zihao Yue assigned | 82 | 3 | 100 OpenMMEgo: Enhancing Egocentric Understanding for LMMs with Open Weights and Data (2025) Video and Temporal Understanding 100 Movie101v2: Improved Movie Narration Benchmark (2025) Video and Temporal Understanding |
|
| 2 | Haiyang Xu 0001 assigned | 65 | 3 | 67 mPLUG-Owl3: Towards Long Image-Sequence Understanding in Multi-Modal Large Language Models (2025) Video and Temporal Understanding 67 TinyChart: Efficient Chart Understanding with Program-of-Thoughts Learning and Visual Token Merging (2024) Multimodal Evaluation and Benchmarks |
|
| 3 | Linchao Zhu | 56 | 3 | 67 From Trial to Triumph: Advancing Long Video Understanding via Visual Context Sample Scaling and Self-Reward Alignment (2025) Video and Temporal Understanding 67 MoS2: Mixture of Scale and Shift Experts for Text-Only Video Captioning (2024) Multimodal Evaluation and Benchmarks |
|
| 5 | Mariano Barone | 52 | 3 | 67 AMICO: A Semantic and Multimodal Framework for AI-Assisted Clinical Reporting (2025) Multimodal Evaluation and Benchmarks 65 A Multilingual Multimodal Medical Examination Dataset for Visual Question Answering in Healthcare (2025) Retrieval-Augmented and Open-Domain QA |
|
| 7 | Tingyu Song | 45 | 3 | 100 VF-Eval: Evaluating Multimodal LLMs for Generating Feedback on AIGC Videos (2025) Video and Temporal Understanding 53 Rethinking Composed Image Retrieval Evaluation: A Fine-Grained Benchmark from Image Editing (2026) Universal Multimodal Embedding and MLLM Retrieval |
|
| 8 | Nancy F. Chen assigned | 45 | 3 | 71 Multimodal Dialogue State Tracking (2022) Conversational Search and Retrieval 67 Topic-Aware Pointer-Generator Networks for Summarizing Spoken Conversations (2019) Multimodal Evaluation and Benchmarks |
|
| 9 | Weipeng Chen | 43 | 3 | 100 Needle In A Video Haystack: A Scalable Synthetic Evaluator for Video MLLMs (2025) Video and Temporal Understanding 67 AITQE: An Adaptive Image-Text Quality Enhancer for Scalable MLLM Pretraining (2026) Multimodal Evaluation and Benchmarks |
|
| 10 | Junjie Zhou 0001 | 43 | 3 | 100 MLVU: Benchmarking Multi-task Long Video Understanding (2025) Video and Temporal Understanding 67 Video-XL: Extra-Long Vision Language Model for Hour-Scale Video Understanding (2025) Video and Temporal Understanding |
|
| 11 | Qingyi Si | 43 | 3 | 67 Towards Flexible Evaluation for Generative Visual Question Answering (2024) Multimodal Evaluation and Benchmarks 67 Multimodal Table Understanding (2024) Multimodal Evaluation and Benchmarks |
|
| 12 | Minghui Fang 0002 | 41 | 3 | 43 Mitigating Hallucinations in LM-Based TTS Models via Distribution Alignment Using GFlowNets (2025) Multimodal Retrieval and Embedding 43 Speech Token Prediction via Compressed-to-fine Language Modeling for Speech Generation (2025) Multimodal Retrieval and Embedding |
|
| 13 | Liwei Wang 0009 | 40 | 3 | 67 AIM: Adaptive Inference of Multi-Modal LLMs via Token Merging and Pruning (2025) Video and Temporal Understanding 42 RAST: Domain-Robust Dialogue Rewriting as Sequence Tagging (2021) Conversational Search and Retrieval |
|
| 14 | Di Wu 0054 | 40 | 3 | 57 The Factuality Tax of Diversity-Intervened Text-to-Image Generation: Benchmark and Fact-Augmented Intervention (2024) Multimodal Evaluation and Benchmarks 43 VisRet: Visualization Improves Knowledge-Intensive Text-to-Image Retrieval (2026) Multimodal and Graph RAG |
|
| 15 | Yanzhao Zhang | 40 | 3 | 67 An End-to-End Model for Photo-Sharing Multi-Modal Dialogue Generation (2025) Multimodal Evaluation and Benchmarks 41 Breaking the Modality Barrier: Universal Embedding Learning with Multimodal LLMs (2025) Universal Multimodal Embedding and MLLM Retrieval |
|
| 17 | Ruichao Yang | 37 | 3 | 64 ExplainHM++: Explainable Harmful Meme Detection With Retrieval-Augmented Debate Between Large Multimodal Models (2026) Multimodal Fake News Detection 64 Towards Explainable Harmful Meme Detection through Multimodal Debate between Large Language Models (2024) Multimodal Fake News Detection |
|
| 18 | Haokun Wen | 37 | 3 | 58 Differential-Perceptive and Retrieval-Augmented MLLM for Change Captioning (2024) Multimodal and Graph RAG 36 Fine-grained Textual Inversion Network for Zero-Shot Composed Image Retrieval (2024) Universal Multimodal Embedding and MLLM Retrieval |
|
| 20 | Jian Lang | 37 | 3 | 41 REDEEMing Modality Information Loss: Retrieval-Guided Conditional Generation for Severely Modality Missing Learning (2025) Universal Multimodal Embedding and MLLM Retrieval 36 Echoes in the Feed: Evolution-aware Prompt-augmented Micro-video Popularity Prediction (2025) Universal Multimodal Embedding and MLLM Retrieval |
|
| 21 | Xinying Qian | 37 | 3 | 36 Hyperbolic Multimodal Generative Representation Learning for Generalized Zero-Shot Multimodal Information Extraction (2026) Universal Multimodal Embedding and MLLM Retrieval 36 SMIR: Span-based multi-grained information refinement for joint multimodal entity-relation extraction (2026) Universal Multimodal Embedding and MLLM Retrieval |
|
| 23 | Jianfeng Dong | 36 | 3 | 43 IVCR-200K: A Large-Scale Multi-turn Dialogue Benchmark for Interactive Video Corpus Retrieval (2025) Conversational Search and Retrieval 39 Rethinking Video Sentence Grounding From a Tracking Perspective With Memory Network and Masked Attention (2024) Video and Moment Retrieval |
|
| 24 | Chen Zhao 0013 | 36 | 3 | 67 Are Multimodal LLMs Robust Against Adversarial Perturbations? RoMMath: A Systematic Evaluation on Multimodal Math Reasoning (2025) Multimodal Evaluation and Benchmarks 36 LimRank: Less is More for Reasoning-Intensive Information Reranking (2025) LLM-Based Reranking |
|
| 25 | Qinggang Zhang | 36 | 3 | 67 Augmenting Intra-Modal Understanding in MLLMs for Robust Multimodal Keyphrase Generation (2026) Multimodal Evaluation and Benchmarks 39 Facilitating Generative Retrieval with Logical Denoising for Interpretable Conversational Search (2026) Personalized and Conversational RAG |
|
| 26 | Idan Szpektor | 35 | 3 | 67 Bridging the Visual Gap: Fine-Tuning Multimodal Models with Knowledge-Adapted Captions (2025) Multimodal Evaluation and Benchmarks 39 MDCure: A Scalable Pipeline for Multi-Document Instruction-Following (2025) Multi-hop and Complex Reasoning QA |
|
| 27 | Sheng-Chieh Lin | 35 | 3 | 53 Unifying Multimodal Retrieval via Document Screenshot Embedding (2024) Universal Multimodal Embedding and MLLM Retrieval 41 Mm-Embed: Universal Multimodal Retrieval with Multimodal LLMS (2025) Universal Multimodal Embedding and MLLM Retrieval |
|
| 28 | Navve Wasserman | 34 | 3 | 41 DocReRank: Single-Page Hard Negative Query Generation for Training Multi-Modal RAG Rerankers (2025) Universal Multimodal Embedding and MLLM Retrieval 41 REAL-MM-RAG: A Real-World Multi-Modal Retrieval Benchmark (2025) Universal Multimodal Embedding and MLLM Retrieval |
|
| 29 | Weiwei Sun 0001 | 34 | 3 | 42 DukeNet: A Dual Knowledge Interaction Network for Knowledge-Grounded Conversation (2020) Conversational Search and Retrieval 34 How Large Language Models Encode Context Knowledge? A Layer-Wise Probing Study (2024) Language Model Pre-training and Adaptation |
|
| 30 | Xuhui Sui | 34 | 3 | 36 Coarse-to-fine multimodal prototype network for few-shot multimodal relation extraction (2025) Universal Multimodal Embedding and MLLM Retrieval 36 SMIR: Span-based multi-grained information refinement for joint multimodal entity-relation extraction (2026) Universal Multimodal Embedding and MLLM Retrieval |
|
| 31 | Xi Ye 0003 | 34 | 3 | 36 Query-Focused Retrieval Heads Improve Long-Context Reasoning and Re-ranking (2025) Multi-hop and Complex Reasoning QA 34 LoFiT: Localized Fine-tuning on LLM Representations (2024) Language Model Pre-training and Adaptation |
|
| 32 | Huiqiang Jiang | 33 | 3 | 42 SeCom: On Memory Construction and Retrieval for Personalized Conversational Agents (2025) Conversational Search and Retrieval 36 Attentive Mask CLIP (2023) Universal Multimodal Embedding and MLLM Retrieval |
|
| 33 | Peitian Zhang | 33 | 3 | 67 Video-XL: Extra-Long Vision Language Model for Hour-Scale Video Understanding (2025) Video and Temporal Understanding 36 INTERS: Unlocking the Power of Large Language Models in Search with Instruction Tuning (2024) LLM-Based Reranking |
|
| 34 | Huaiwen Zhang | 33 | 3 | 68 Evaluating and Mitigating Sycophancy in Large Vision-Language Models (2025) Multimodal Evaluation and Benchmarks 39 Debiased Video-Text Retrieval via Soft Positive Sample Calibration (2023) Video and Moment Retrieval |
|
| 35 | Archiki Prasad | 33 | 3 | 44 GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs (2026) Language Model Pre-training and Adaptation 43 How Accents Confound: Probing for Accent Information in End-to-End Speech Recognition Systems (2020) Multimodal Retrieval and Embedding |