multimodal system developers

Abstract: As multimodal agents accumulate experience, their long-term memories contain many similar observations of recurring scenes, actions, and entities. Answer quality depends on the evidence retrieved from these memories into the model's context. However, query-relevant memories may contain overlapping information, and a fixed retrieval depth may return too little or too much evidence for a given question. To address these challenges, we propose PACE-Mem, Prefix-Aware Complementary Evidence, which constructs a complementary evidence sequence and determines its length from the same selection trajectory. For structured memories, query-guided pooling combines records within each clip to obtain relevance scores and representations. Using these relevance scores and representations, maximal marginal relevance (MMR) selection balances query relevance against overlap with previously selected evidence. From the resulting score trajectory, a prefix-length selection algorithm determines retrieval depth by comparing length-normalized adjacent declines, accounting for both score changes and the amount of evidence accumulated. PACE-Mem is a training-free, plug-in retrieval method for clip-based agent memory systems, leaving memory construction and answer generation unchanged. Across three long-video benchmarks and two agent frameworks, PACE-Mem improves question-answering accuracy by an average of 1.6% over Adaptive-k, the strongest baseline at comparable mean retrieval depths. Further analysis reveals that jointly adding and removing evidence is associated with larger net accuracy gains than either change alone.

Abstract: Multimodal sentiment models are usually evaluated on complete utterances, yet streams often lose a contiguous span of text, audio, or video. We study local missingness on aligned CMU-MOSEI features. Our reliability-aware network (RM-Net) keeps the original temporal positions, represents unobserved positions with a learned token, outputs bounded per-modality intensity contributions gated by observation quality, and is trained with dual-view masking of contiguous segments. We fuse five RM-Net seeds with a LoRA-adapted BERT text expert, using a fusion weight that depends only on the visible fraction of content tokens. Architecture, checkpoint, and fusion choices were made on the validation split. On the held-out split, the fused model reaches a macro-F1 of 0.655 on clean inputs and 0.634 under 40\% mixed missingness; its two-view average exceeds the auxiliary-head baseline by 0.024 (95\% video-cluster bootstrap interval [0.003, 0.045]). Because this split was inspected during development, these numbers are not a blind test. For explanations, exact Shapley attributions over the eight modality subsets and position-level occlusion show that top-ranked positions reduce the predicted-class probability faster than random positions (AUC gap 0.113 on validation). We also describe an auditable word-level alignment pipeline for raw clips; its outputs do not feed the predictive models.

Abstract: Vision Transformers incur high computational costs when processing long visual token sequences. Existing reduction methods typically rank tokens independently and discard those deemed unimportant or redundant. We propose a reconstruction-guided framework for visual token condensation (ReConV), which summarizes the complete visual sequence with a compact Summary Bank (SMB) and retains tokens whose information is not adequately captured by the SMB, as indicated by their reconstruction residuals. At each compression layer, a lightweight position-conditioned decoder reconstructs the current tokens, leaving shared information compressed in the SMB while explicitly preserving poorly reconstructed details. Experiments on image classification and multimodal understanding demonstrate favorable performance-efficiency trade-offs: ReConV surpasses the uncompressed baseline with 48\% fewer GMACs on ImageNet and enables reusable image-level representations for efficient multi-turn inference. Code is available at TokenDrop.

Abstract: A vision-language model may need to combine an image with a prompt to recognize a safety risk that neither reveals alone. Where does this joint safety judgment become accessible inside the model? We introduce SSU-Bench, a dataset of matched safe and unsafe image-text combinations constructed using single-item prompt edits or image edits with annotated intended regions. Using three vision-language models, we transfer internal states between paired inputs and measure the resulting change in the safety verdict. Across models and both types of counterfactual, interventions at the changed input positions are effective in earlier decoder layers, while interventions at the final input token become effective later. Directions estimated from other examples produce similar late-layer effects. A linear readout of the final-token state also predicts the model's own verdict, including incorrect judgments, and cross-model comparisons reveal similarities in the patterns of counterfactual change. These findings identify a recurring transition in where interventions can influence a joint safety verdict and distinguish a readable model decision from a correct safety judgment.

Abstract: CLIP-like vision-language models remain a cornerstone of multimodal systems, yet their scores stay near chance on directed spatial relations, such as whether one object is left of another. We call this failure readout blindness and analyze, theoretically and empirically, why deployed scores miss the direction: when scoring rules treat the subject and object symmetrically, direction cancels regardless of encoder training. Guided by this analysis, we introduce Antisymmetric Displacement Readout (ADR), which aligns caption words with image patches in the frozen features and scores each relation by the signed displacement between matched object centroids. Notably, ADR succeeds without additional training or learned parameters, thereby demonstrating that directional information remains in the frozen encoder. However, text and world priors can inflate accuracy, so we further introduce prior deflation, which measures the benefit of the image-text pairing as the grounded gain over a null that pairs each item with an unrelated image. Extensive experiments across encoder families show that ADR substantially improves over deployed scores, which remain near chance on most direction-balanced sets even for fine-tuned encoders. Compared with more complex readouts, ADR outperforms the evaluated MLLM likelihood readouts and is competitive with their chat inference at a small fraction of the computation. These results support our claim that directional information can be recovered from frozen features by an appropriate readout. Our implementation and evaluation kit will be publicly available.

Abstract: The ability of vision-language models (VLMs) to associate visual identities with biographical information creates a need for selective unlearning of personally identifiable information (PII) while preserving permitted knowledge about the same individual. This setting is challenging because both sensitive and retained information can share the same visual inputs and intermediate representations. We introduce SIEVE, a simple and effective framework for selective VLM unlearning. SIEVE directly regularizes attention-value representations while also controlling model outputs. SIEVE suppresses attention values for forget examples toward a constant zero, while preserving retain-example representations by matching them to a frozen reference model. These objectives are combined with sequence-level forget and retain supervision, enabling targeted forgetting without largely affecting retained knowledge. Extensive experiments show that SIEVE achieves state-of-the-art performance on unlearning with multiple model-modality settings, while maintaining competitive retained utility. Ablation studies further show that value suppression and negative cross-entropy contribute complementary forgetting signals, while reference-based value matching substantially reduces utility degradation. These results demonstrate that attention values provide an effective intervention point for selective multimodal unlearning when sensitive and retained knowledge are closely related.

Abstract: Adaptive RAG often uses retrieval-time signals to decide whether another retrieval, reranking, or multimodal step should run. We ask whether these signals add routing value once the query itself is already known. Across document, audio, and video RAG, we compare matched query-only and query+retrieval routers while holding the optional actions, router family, training procedure, and evaluation fixed. On the held-out final evaluation, adding the tested retrieval signals does not produce a reliable routing improvement over the query-only baseline. Some retrieval signals are associated with whether a later step will help, but that predictability does not consistently lead to bet- ter RUN/SKIP decisions. The main lesson is therefore methodological: retrieval-state features should not be credited with routing value unless they improve over a matched query-only control. Our results do not show that routing or retrieval state is generally useless; they show that the incremental value of retrieval signals must be demonstrated rather than assumed.