The resulting training-free pipeline, RECAST, consistently outperforms recent frame-selection baselines without modifying the ITM encoder, and preserves the per-frame matching cost of a standard single-query baseline.
Abstract
Long-video question answering (QA) forces multimodal large language models (MLLMs) to work within a tight frame budget, so the choice of frames largely decides whether a question can be answered at all. The standard recipe scores every frame against the question with a pretrained image–text matching (ITM) model and keeps the top scorers. A fundamental mismatch underlies this recipe: ITM models are trained on short, concrete visual descriptions, while QA questions are interrogative and often involve abstract terms. Scored against the question alone, the ITM yields a near-random signal whenever the question is not a direct image–text match, such as one asking for the temporal order of scenes. In our LongVideoBench diagnostic analysis, the score collapses even on benchmark-provided answer-relevant frames, with more than half falling into a near-zero region —not because the encoder is faulty, but because it behaves exactly as it was trained to. We argue that this format mismatch should be absorbed at the two ends of the pipeline while the encoder itself stays frozen. On the input side, a type-conditioned routed pipeline reformulates each question into a single ITM-aligned description by selectively applying grounding, decomposition, and constrained synthesis. The ITM therefore receives exactly one description per frame, preserving the per-frame matching cost of a standard single-query baseline. On the output side, because the score distribution remains polarized and answer frames are scattered in time, we replace top- $K$ selection with a parameter-free Rosin threshold followed by a temporal maximal-marginal-relevance (MMR) step that uses frame positions alone. Across three benchmarks (LongVideoBench, Video-MME, MLVU) and and four backbones (Qwen2-VL, Qwen2.5-VL, LLaVA-OneVision, LLaVA-Video), the resulting training-free pipeline, RECAST, consistently outperforms recent frame-selection baselines without modifying the ITM encoder.
Q-Guide is built, a small agent that reads a question, works out what evidence it is still missing, and calls targeted tool(s) to recover it---reading text where text is needed, zooming in where detail is needed, or grounding a region where position matters.
DCASE~2026 Task~5 introduces Audio-Dependent Question Answering (ADQA), which tests whether large audio-language models answer from the audio rather than from textual priors. An Audio-Dependency Filtering (ADF) pipeline combines silent-audio probing, per-option perplexity, a large language model (LLM) commonsense check, and human review to remove items solvable from text alone. The 3000 items that pass form the ADQA-Bench evaluation set, spanning music, speech, and environmental audio. The inaugural edition draws 14 teams and 36 submissions across two tracks defined by total parameter count (up to 100B and under 10B). A Chung-Ang University ensemble of MOSS-Audio-8B-Thinking and Qwen3-Omni-30B reaches the top overall accuracy at \pct{58.33}, and a MOSS-only configuration from the same team leads the sub-10B track at \pct{57.30}. Across the 30 submissions with a comparable development score, evaluation accuracy falls by 11.91 percentage points (pp) on average (median 10.91\,pp) on the hidden evaluation split, which is designed to be harder than the development split. The most common building blocks are: the MOSS-Audio-8B-Thinking backbone (13 of 36 submissions), Low-Rank Adaptation (LoRA) fine-tuning on AudioMCQ-StrongAC, and preference or reinforcement-learning objectives -- Group Relative Policy Optimization (GRPO) in five teams, Group reward-Decoupled Normalization Policy Optimization (GDPO) in two. At test time, prompt engineering is near-universal, and majority or choice-permutation voting is common. Every system misses the same set of 233 evaluation items.
Haolin He, Renhe Sun, Zheqi Dai et al.· 1 citation
Conversational Image Retrieval (CIR) refines image search through multi-turn interaction, where the Questioner plays a central role in eliciting information about the user’s target. However, existing CIR systems are commonly evaluated in end-to-end settings, making it difficult to determine whether performance gains originate from questioning strategies, retrieval backbones, answering behavior, or interaction protocols. We introduce a controlled benchmark for studying questioning strategies in CIR. The proposed benchmark isolates the Questioner module by keeping the retriever, answerer, dialogue budget, decoding setting, and evaluation pipeline fixed across all methods. Under this unified protocol, we compare representative strategies including Blind, Text-based, Text-based Recon, Top-K Guided, and Hybrid. Experimental analysis reveals a clear stage-dependent trade-off between efficiency and discriminative capability. Text-based strategies remain computationally lightweight but become increasingly vulnerable to contextual drift over extended interaction, whereas Top-K Guided improves retrieval refinement at the cost of higher inference latency. Overall, the proposed benchmark provides a reproducible framework for analyzing conversational questioning behavior in CIR systems.
Bui Tay, Son Nguyen Thanh, Phuc Nguyen Vu et al.· International Conference on...· 0 citations
Streaming automatic speech recognition (ASR) must be judged jointly on what it transcribes and on how quickly it commits each word. Delayed streams modeling (DSM) has become the dominant paradigm for streaming large audio-language models, exposing a structural delay $\tau$ that bounds the decoder's lookahead. We show that $\tau$ is a poor proxy for user-perceived latency, and that the alignment-based supervision of DSM leaves latency on the table: the same forced-aligned transcript is used at every $\tau$, forcing the model to withhold words it could already commit. We introduce AWED, a word-level emission-delay metric defined relative to the acoustic end of each word, and post-train a DSM recognizer with GRPO under a reward that scores transcription accuracy and measured delay jointly. Trained at a single operating point ($\tau=6$ frames), our model dominates both its supervised fine-tuning initialization and the Voxtral Realtime backbone across all evaluated lookahead budgets: it cuts WER by 30.8\% relative at an 80\,ms structural delay, and by 5.7\% relative at 480\,ms while lowering median AWED from 1.17\,s to 1.04\,s. Latency-rewarded post-training thus advances the accuracy--latency Pareto frontier of streaming ASR without architectural change.
Large Audio-Language Models (LALMs) have recently shown strong capabilities in speech understanding and question answering (QA), but they also inherit privacy risks from large-scale training data, including the unintended memorization of sensitive information. In this work, we study machine unlearning for speech QA in LALMs, a setting that is more challenging than prior work on text-based Large Language Models (LLMs) or Automatic Speech Recognition (ASR) due to the tight coupling between acoustic perception and factual knowledge. We present and evaluate multiple unlearning strategies, including gradient ascent, task arithmetic, and alignment-based fine-tuning methods that enforce safe refusal responses, to remove private knowledge while still preserving performance on core capabilities. Through extensive experiments on speech QA datasets, we show that these unlearning methods can reduce the privacy leakage rate by up to 80% while maintaining near-neutral performance on non-private speech QA and general speech understanding benchmarks.
This work presents MMOOC, a large-scale benchmark for evaluating refusal and robust answering abilities of MLLMs, and introduces an LLM-as-a-Judge metric to assess the correctness of model reasoning.
