Attention
-
BAT3R: Robust Online 3D Reconstruction with Bayesian Adaptive State Updates
2026-10-12 07:47:47Lijian Li, Yuanpeng He, Yudian Zheng, Chung-ju Huang, Linyu Li, Dongming Jin, Wenpin Jiao, Zhi Jin, Chi-Man PunAbstract
Feed-forward transformer models have driven remarkable progress in 3D vision,yet their quadratic complexity renders them impractical for long video sequences. Streaming 3D reconstruction addresses this bottleneck by processing frames sequen?tially with constant memory. However, existing recurrent architectures inevitably suffer from progressive degradation over long sequences, as they fail to account for the intrinsic reliability of observations and the accumulated confidence of the historical state, leading to severe error accumulation. To address these issues, we introduce BAT3R, a training-free Bayesian adaptation framework comprising two synergistic modules: State Innovation Gating (SIG) and Bayesian-guided Adaptive State Updating (BASU). SIG acts as a distributional regularizer that anchors state evolution to its geometric initialization, effectively arresting representational drift during long-term inference. Building on this, BASU reformulates state update as a recursive Bayesian inference process within a Kalman filtering framework. By jointly modeling the temporal reliability of observations via information-theoretic gain and spatial relevance via cross-attention maps, BASU achieves an asymmetric spatio-temporal state fusion: well-consolidated historical regions are protected from redundant updates, while genuinely novel scene content is aggressively incor?porated. Extensive experiments on standard benchmarks (7-Scenes, TUM, ScanNet, KITTI, etc.) demonstrate that BAT3R consistently outperforms state-of-the-art online 3D reconstruction methods, with the most significant gains observed on long sequences and complex scenes.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16330
PDF
-
Explicitly Aligned Multi-View Breast Cancer Risk Prediction from Variable-Length Screening Histories
2026-10-12 07:47:42Solveig Thrun, Zijun Sun, Suaiba A. Salahuddin, Kristoffer Wickstr{\o}m, Elisabeth Wetzer, Stine Hansen, Robert Jenssen, Michael KampffmeyerAbstract
Accurate breast cancer risk prediction from screening mammography is essential for improved risk stratification and personalized screening. It can benefit from both temporal breast tissue changes across longitudinal examinations and complementary anatomical information from craniocaudal (CC) and mediolateral oblique (MLO) views. Leveraging longitudinal information requires identifying corresponding tissue regions across examinations, yet most approaches rely on implicit feature alignment without explicitly enforcing spatial consistency. Recent work has demonstrated the benefits of explicit temporal alignment; however, existing methods model each view independently, preventing them from exploiting complementary CC and MLO information. Furthermore, explicit approaches have been limited to a single prior examination. To fill these gaps, we propose (this http URL), a breast-wise longitudinal multi-view framework that jointly models CC and MLO views across screening examinations while accommodating a variable number of prior examinations. FlexLMV-Net explicitly aligns longitudinal feature maps and integrates complementary views through cross-view temporal fusion and a dual-stream attention block. Evaluated on EMBED and CSAW-CC, FlexLMV-Net consistently outperforms existing approaches across breast density, cancer subtypes, and demographic subgroups, highlighting the benefits of explicit longitudinal alignment, multi-view modeling, and flexible use of available screening history. We further introduce (this http URL), an open-source benchmark that unifies state-of-the-art single- and multi-timepoint breast cancer risk prediction models under standardized preprocessing, training, inference, and evaluation protocols, enabling fair and reproducible comparison of existing methods.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16329
PDF
-
FON: A Draft-Revise Dual-Organ Language Model without Sequence-Level Quadratic Attention
2026-10-12 07:28:24Helong Hu (Zhanjiang Campus, Guangdong Ocean University, Zhanjiang, Guangdong, China), Hongdan Pan (Maoming Administrative Service Center, Maoming, Guangdong, China), Shuiqing Hu (China Pacific Insurance)Abstract
Self-attention assigns two functions to one mechanism-where to read and how to mix-at the price of O(L^2) training, an O(T) decode state, and position signals bound to the training length. Existing attention-free alternatives all optimize the single function of generation. We propose FON (Fusion-Organ Network), which differentiates one weight stack into a drafting organ (multi-tap learnable delays + content-adaptive recursion; causal, O(S) training, O(1) decode state, no sequence-level quadratic attention) and a revision organ that re-processes the written prefix through multi-scale ripple kernels injected via weak gates, never generating. At 30M parameters, iso-parameter and iso-token budget, FON reaches perplexity 131.03 on WikiText-103, beating Transformer (197.73) and Mamba2 (152.53); the advantage persists under all depth-aligned baselines and at the 70M tier (-20.8%). At 8x length extrapolation FON degrades by ~8-9% without positional-encoding engineering, between ALiBi and RoPE. The revision channel provides a structural pathway for in-domain infilling that causal architectures lack in principle (+19% over on-demand coupling), at a perplexity cost indistinguishable within measurement precision. We report all costs and boundaries: impaired long-range retrieval, synthetic-task disadvantages, weaker out-of-domain code infilling than a same-recipe Transformer, and a 6.5-12.3x wall-clock gap.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16301
PDF
-
MP-MNCA: Morphology-Prior Masked Neighbor Cross-Attention for Efficient Spatial Domain Identification
2026-10-12 05:42:01Sidharthan Sivacoumar Chitra, Sanjeev ManivannanAbstract
Spatial transcriptomics jointly captures molecular state, tissue morphology, and spatial organization, yet spatial-domain methods are typically compared only at late training checkpoints, obscuring differences in convergence efficiency. We introduce MP-MNCA, a morphology-prior masked neighbor cross-attention framework that preserves the full 3,000 dimensional highly-variable-gene representation while selectively routing molecular information across spatial neighbors, with histological similarity incorporated as a continuous attention prior under a self-supervised masked neighbor-contrastive objective. Across 12 heterogeneous human dorsolateral prefrontal cortex (DLPFC) sections, MP-MNCA is evaluated against established spatial-domain baselines using both convergence trajectories and epoch-wise spatial-domain visualizations. Our results show that MP-MNCA (i) reaches coherent spatial-domain structure in substantially fewer training epochs, as reflected quantitatively by convergence curves and qualitatively by progressive spatial grids, and (ii) maintains strong clustering quality, achieving a mean peak refined ARI of 0.546 versus 0.549 for the strongest baseline (STAIG), while attaining the highest mean refined NMI of 0.657 versus 0.640. Moreover, MP-MNCA outperforms STAIG on 7 of 12 sections in ARI and 11 of 12 in NMI, demonstrating faster convergence without sacrificing spatial-domain quality.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16212
PDF
-
Omni-Modal LLMs Know More Than They Say
2026-10-12 03:52:46Na Min An, Taekyung Kim, Jaemyung Yu, Hyunjung Shim, Dongyoon HanAbstract
Omni-modal large language models (Omni-LLMs) extend LLMs to process heterogeneous modalities (e.g., vision, audio, and IMU signals) alongside language, enabling a broader range of multimodal tasks. Yet we find that these models often fail to express information from non-textual modalities even when they perceive it correctly. We term this disconnect knows-but-cannot-say. To investigate this phenomenon, we introduce Token-level Modality Alignment (TMA), which measures the alignment between individual modality representations and response representations. Our analysis shows that non-textual representations are less aligned with response tokens and receive less attention than textual representations, suggesting a potential explanation for uneven modality preferences. Motivated by these findings, we propose OmniCatalyst, a simple perceive $\rightarrow$ infer strategy that complements the original modality inputs with self-generated text, making non-textual knowledge more accessible. Across seven Omni-LLMs, OmniCatalyst consistently improves performance (+7.29%) and yields more balanced modality preferences, with further gains from diverse training strategies. Further analysis demonstrates that the knowledge extraction phases increase TMA, suggesting that improved alignment makes non-textual information more accessible during response generation. In summary, our findings uncover how Omni-LLMs can better express modality knowledge through language. Our code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16135
PDF
-
OT-FlowRec: Optimal Transport-Guided Multi-Bridge Flow for Multimodal Sequential Recommendation
2026-10-12 01:23:07Miao SunAbstract
E-commerce user preferences evolve dynamically within sessions, exhibiting concurrent and multi-centric interest distributions. While existing multimodal sequential recommendation methods leverage global attention to compress historical interactions, they inadequately capture the concurrent nature of multi-interest behaviors and fail to model the underlying multi-centric distribution of user intent. To tackle this issue, we propose OT-FlowRec, an optimal transport-guided multi-bridge flow matching framework that explicitly models multi-interest distributions to facilitate multimodal sequential recommendation. Unlike single-path generative models, OT-FlowRec constructs a mixture of parallel, interest-specific transport flows to model the multi-centric distribution characteristics, where each flow preserves a distinct user interest cluster via an optimal transport plan. Then, a geometric transport mechanism is designed to establish minimum-cost paths from the latent multi-interest distribution to target preference prototypes to provide multi-interest awareness. Moreover, a progressive time-sampling warmup strategy dynamically biases optimization weights toward the final states of the generation process, bridging the gap between continuous global distribution fitting and practical top-$k$ ranking metrics. This approach achieves explicit interest disentanglement, ensuring the preservation of a multi-peaked preference representation instead of collapsing it into a unimodal average. Extensive experiments on three public datasets demonstrate that OT-FlowRec outperforms existing state-of-the-art baselines.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16038
PDF
-
No Concept Escapes the Audit: Auditing-Aware Unlearning for Verifiable Concept Erasure in Diffusion Models
2026-10-11 20:57:56Kaiyuan Deng, Yuchen Li, Gen Li, Yang Xiao, Geng Yuan, Xiaoyong Yuan, Bo Hui, Xiaolong MaAbstract
Text-to-image diffusion models can generate prohibited content, which motivates concept erasure through machine unlearning. Most erasure methods intervene at the text interface, through prompt modification or localized updates to text-conditioning weights, and they are evaluated by what the model outputs for given prompts. Such evaluation cannot see what the network still encodes. Latent-space auditing, which bypasses text conditioning and probes the denoising network directly, shows that erased concepts remain recoverable from internal representations. We find that this also holds for methods built to be robust against adversarial prompts, and that the problem grows with the number of erased concepts. We propose Auditing-Aware Unlearning for Verifiable Concept Erasure in Diffusion Models (AVCE), a framework that grounds erasure in the model's latent representations. AVCE audits the embedding neighborhood of each concept and condenses the discovered vulnerable directions into an anchor at the weakest geometric point. It edits cross-attention and self-attention projections in closed form at this anchor, then fine-tunes the two pathways with pathway-level auditing losses, using orthogonal gradient projection to consolidate multiple concepts. Experiments on SD v1.5, SDXL, and Flux 1.0 across object, explicit-content, and artistic-style unlearning show that AVCE reduces attack success rates by 5.07x and improves auditing scores by 3.84x over the strongest baseline, while preserving competitive generation quality.
Abstract (translated)
URL
https://arxiv.org/abs/2610.05401
PDF
-
What the Model Cannot See: Covert and Provable Text Invisibility in LLMs
2026-10-11 20:00:37Mark RussinovichAbstract
Text wrapped in a trigger can be made behaviorally invisible to a language model: omitted from summaries and quotations, ignored in question answering, and left unflagged by downstream filtering. We study this selective blindness as a dual-use information-flow primitive, a supply-chain backdoor, and a white-box auditing problem. We first implement it exactly as a serving-time span veto: blocking information flow out of a marked span makes every outside activation bit-identical under same-length substitution of the span's content, a noninterference property we verify across ten models, seven mixer families, and 0.5B-120B parameters. The veto serves as a deployable quarantine, an exact reference for calibrating leakage, and a teacher. Distilled into weights, it yields a backdoor: checkpoints trained on six model families from 4B to 120B ignore marked context at little benchmark cost, and their outputs show no measurable semantic trace of the suppressed content above the veto reference on our held-out suite; an in-domain variant suppresses planted vulnerabilities during AI code review. Detection is therefore a white-box problem. Simpler backdoors remain internally transparent -- linear probes recover suppressed content at 0.57-0.79 accuracy versus 0.025 chance -- but hardening the attack until probes reach chance collapses attention from outside the span into it below the range of base models. Across the attacks we evaluate, probe visibility trades off against routing visibility; we do not establish that this tradeoff is fundamental.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15883
PDF
-
Attention Sinks in VLMs Are Borrowed, Not Seen
2026-10-11 19:20:41Dengjia Zhang, Alexander Martin, Pengyu Yan, Xihan Sun, Jiaqi Li, Benjamin Van Durme, Andrew Yates, Reno Kriz, William WaldenAbstract
We study how to detect and mitigate visual attention sinks in vision-language models (VLMs). Visual attention sinks are image tokens that absorb a large share of attention over the image regardless of the query or task, and they mislead methods that treat attention maps as a record of where the model ``looked.'' Prior work has detected sinks through massive activation dimensions and proposed redistributing their attention, but none have addressed what makes a token become a sink. Comparing six candidate detectors on ten VLMs from four families, we find that only the centered cosine similarity between an image token's key and the key of the language model's text-sink detects visual sinks. Thus, a visual sink is an image token whose key aligns with the text-sink's key. Moving a sink's key component along this direction to the median of the image keys sends its attention back to the text-sink (\textbf{Are Borrowed}) while leaving model answers unchanged~(\textbf{Not Seen}). Building on this, we propose a training-free, entropy-triggered re-look that first cleans the attention map, then crops the most-attended region of the cleaned map and inserts it into the reasoning trace. Cleaning adds negligible computation to the forward pass, and the re-look improves accuracy with a mean gain of $+4.3$ points on fine-detail VQA benchmarks across model families.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15855
PDF
-
An Explainable System for Offensive Content Detection in Nastaliq Urdu: Corpus Curation, Transformer Benchmarking, and Word-Level Attribution
2026-10-11 17:33:09Muhammad Shahid Iqbal MalikAbstract
Automated moderation systems are deployed at scale, yet they remain opaque to the users and moderators affected by their decisions, and their coverage is almost entirely confined to English. For Urdu written in the Nastaliq script, no such system exists, and the annotated resources it would be built on are drawn largely from Twitter/X, where character limits constrain how users write. Interpretability has received little attention, leaving it unclear what such models actually learn. This study addresses these gaps in three steps. First, we curate and publicly release a corpus of 7604 Nastaliq Urdu Facebook comments, extending an earlier corpus by correcting text-level inconsistencies and re-examining the original labels. Second, we fine-tune four Urdu-capable transformers under a common grid-search protocol and compare them against twelve feature-classifier baselines on identical data partitions. The best configuration, Multilingual Representations for Indian Languages (MuRIL), reaches 87.05% accuracy and 87.06% macro F1, improving on the strongest baseline by 5.52 percentage points, a margin that exceeds the 95% confidence interval on test accuracy; the two leading transformers differ by less than the test set can resolve, and performance is stable across five optimizers. Third, we compute word-level SHAP attributions adapted to right-to-left rendering. Attribution mass concentrates on high-frequency slurs but extends to syntactically neutral function words and to political vocabulary characteristic of the source pages, indicating that aggregate performance may rest partly on lexical and topical regularities rather than a generalizable representation of offensiveness.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15795
PDF
-
History Replays Itself: Unleashing Long-Context Capability in Linear Attention
2026-10-11 15:58:04Li Nanbo, Firas Laakom, Yilan Zhang, Piotr Pi\k{e}kos, Yimeng Chen, Liangyu Wang, Yifan Xu, Mateusz Ostaszewski, Dandan Guo, Wenyi Wang, Yanning Dai, Yuhui Wang, J\"{u}rgen SchmidhuberAbstract
Linear attention promises the efficiency needed for long-context language modelling, yet its fixed-size recurrent state must simultaneously support ongoing computation and preserve information from an ever-growing history. This coupling has reinforced a prevailing view that the state's memory capacity defines the model's long-context capability. We challenge this presumed ceiling: \emph{recurrent computation need not bear the full burden of long-term storage}. We introduce NoteBank, a sparse episodic memory that archives selected text before future queries are known, uses the backbone's native features for writing and addressing, and reads retrieved evidence through its native forward computation via replay. This \emph{training-free, plug-and-play} interface requires no external embedding model, separately trained retriever, or separate neural reader. At one million tokens, it raises Gated DeltaNet's mean accuracy across six RULER retrieval tasks from 1.7% to 70.3%, with no observed significant retrieval degradation from 4K to 1M tokens. This recovery adds $\leqslant$0.025% to analytical FLOPs, with persistent context storage approximately 1/2,600 that of a Transformer's full MHA KV cache at matched width, depth, and precision. Experiments extend from three 1.3B recurrent backbones to Mamba-2-8B and post-trained Qwen3.5-9B, demonstrating gains on LongBench v1/v2 and MemoryAgentBench. Together, these results show that sparse evidence storage and native replay can substantially expand the long-context capability of pretrained linear attention models while retaining the efficiency of recurrent computation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15715
PDF
-
Self-Representation Alignment in Diffusion Models for Video Background Music Generation
2026-10-11 12:40:40Jincheng Zhang, Gy\"orgy Fazekas, Charalampos SaitisAbstract
Diffusion models have been widely used for various generative tasks. However, recent studies have demonstrated that the internal representations learned in the diffusion models' denoising process are insufficiently meaningful, which can hinder the generation performance. To address this, we propose a novel diffusion U-Net architecture enhanced by explicitly aligning the weaker representation from the earlier layer conditioned on higher noise with the better representation from the later layer conditioned on lower noise during training, which we use for the challenging multimodal task of video background music generation. Our diffusion U-Net Self-Representation Alignment (U-SRA) employs an alignment loss to effectively guide our diffusion model to learn good internal representations, thereby enabling it to produce higher-quality symbolic music conditioned on video features. We employ a cross-attention mechanism to synchronize music with video segments. Experiments on a benchmark video-music dataset demonstrate that the proposed U-SRA approach for video background music generation surpasses the state-of-the-art across key evaluation metrics, including video-music alignment and overall music quality.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15575
PDF
-
ELECT: Settling Identity Before Rendering for Scalable Story Visualization
2026-10-11 11:57:18Mehrajul Abadin Miraj, Md. Mahmudul Hasan, Meherun Farzana, Aniket Joarder, Mahmudul Hasan, Abir Chakraborty Partha, Md. Ahasanul Alam, Redwan Ahmed Rizvee, Md. Fahim Arefin, Md Mahmudur Rahman, Md. Tanvir Alam, Md Mosaddek KhanAbstract
Training-free story visualization without reference images must preserve recurring entities across frames while scaling to long narratives. Existing methods typically maintain consistency by conditioning each frame on other frames of the same story, coupling generation time, VRAM, or both to story length. We introduce ELECT, which instead settles cross-frame identity before final rendering through a language-guided scene-entity mapping and stationary anchors elected from the model's own drafts. This makes the per-frame cost of the final rendering pass independent of story length. Joint attention exceeds 96 GB of VRAM at 32 scenes, whereas ELECT remains at 8.3 s per frame and under 20 GB VRAM from 10 to 1000 scenes. On the Story50 long-story benchmark, ELECT reaches 0.833 average Character Consistency Similarity (aCCS), compared with 0.826 for ten-round refinement, at one seventh of its generation cost, with lower text alignment (0.327 vs. 0.340). At 1000 scenes, to our knowledge, the first evaluation of training-free story visualization at this length, ELECT exceeds ten-round refinement by 0.017 aCCS and the nearest flat-cost baseline by 0.091, while also surpassing refinement in text alignment at one twenty-fourth of its generation time. These results show that long-horizon consistency need not require increasingly expensive inter-frame conditioning as stories grow.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15555
PDF
-
Sleep Sound Event Detection Powered by Learnable Multi-Resolution Adaptive Line Enhancer
2026-10-11 11:10:08Chanwoo Park, Chanwoo KimAbstract
Sleep-disordered breathing, including snoring, hypopnea, and obstructive apnea, is the hallmark of obstructive sleep apnea, whose severity is quantified by the Apnea-Hypopnea Index (AHI). As conventional polysomnography is costly and clin- ically demanding, microphone-only detection has emerged as a noninvasive alternative for screening. We propose ALE- Confidence Fusion Sound Event Detection (ACF-SED), the first framework that directly injects the per-frame confidence map of an Adaptive Line Enhancer (ALE) into Transformer atten- tion biasing and feature modulation, rather than as mere pre- processing. A Multi-Resolution ALE Bank aggregates three parallel NLMS filters at distinct decorrelation delays, while a Confidence-Guided Cross-Path block couples event and noise streams via gated cross-attention. On the Audio-Polygraphy Dataset for Sleep Apnea Analysis, ACF-SED achieves state- of-the-art detection across Event-F1, Segment-F1, and PSDS, enabling end-to-end AHI estimation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15532
PDF
-
CacheLink: Functional KV Transport for Zero-Prefill Cross-Model Context Handoff
2026-10-11 10:19:31Haomiao Qiu, Yining Tang, Zijia Song, Ruizi Han, Ziyue Qiao, Miao ZhangAbstract
KV caches make long-context serving efficient by reusing past computation, but model switching breaks this reuse and forces the receiver to prefill the context again. We study \emph{zero-prefill cross-model context handoff}: translating a sender cache into a receiver-compatible state so that the receiver can continue directly without prefilling same context. Existing affine transport methods show that receiver KV states can be substantially reconstructed from sender caches. However, through extensive analyses, we find that reconstruction fidelity is a poor proxy for receiver utility. Similar K/V reconstruction errors can lead to substantially different attention patterns, predictive distributions, and downstream accuracy, because conventional reconstruction objectives overlook the query-dependent functional importance of K and V in receiver attention. We thus introduce CACHELINK, which learns a coupled K/V correction through the receiver behavior directly. Across four independently trained sender--receiver pairs and seven evaluation panels, CACHELINK improves the equal-weight mean over all methods for every pair. In matched diagnostics, CACHELINK changes K/V squared error by less than 0.4% while substantially changing receiver attention and next-token behavior. At 32K context, it reduces handoff latency by up to $7.27\times$ relative to receiver re-prefill. Our findings verify that reconstruction fidelity alone is insufficient and that preserving the receiver computation induced by the transported cache is important for effective cross-model KV transport.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15509
PDF
-
E8 Lattice Key Quantization for Low-Bit KV-Cache Compression
2026-10-11 06:22:28Sasho Nedelkoski, Alexander Acker, Soeren Becker, Dominik Scheinert, Odej KaoAbstract
Compressing the key-value (KV) cache is essential for serving long-context large language models (LLMs) under memory constraints. While current scalar quantization methods achieve significant compression, they typically face a quality-fidelity trade-off, either suffering from performance instability at 2-bit precision or necessitating 3-bit keys to maintain model quality. In this work, we demonstrate that a uniform low-bit k2v2 configuration (2-bit keys and 2-bit values) can preserve near-baseline quality by using vector quantization (VQ) based on the $E_8$ lattice for KV keys. We show that structured VQ exploits correlations across adjacent dimensions to achieve signal-to-quantization-noise ratios (SQNR) higher than those of independent scalar quantization at the same key payload rate. To ensure practical utility, we adapt a table-based asymmetric scoring algorithm that avoids explicit codebook dequantization during attention computation and preserves decode-throughput parity with scalar baselines. Across five LLM families, our $E_8$ k2v2 approach yields \textasciitilde$6\times$ KV compression, roughly 20\% smaller than scalar k3v2 (3-bit keys and 2-bit values), while retaining near-baseline long-context perplexity. In our full-suite perplexity evaluation, $E_8$ improves over scalar k3v2 in 80\% of model--dataset--context settings. These results indicate that $E_8$ lattice VQ improves quality per byte while keeping throughput at parity, suggesting that structured key quantization is a useful primitive for KV-compression stacks. Overall, our work provides a low-overhead primitive for expanding the quality-memory frontier in long-context LLM inference.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15384
PDF
-
Stochastic CHAOS: Why Deterministic Inference Kills, and Distributional Variability Is the Heartbeat of Artifical Cognition
2026-10-11 06:02:40Tanmay Joshi, Shourya Aggarwal, Anusa Saha, Aadi Pandey, Shreyash Dhoot, Vighnesh Rai, Raxit Goswami, Aman Chadha, Vinija Jain, Amitava DasAbstract
Deterministic inference is a comforting ideal in classical software: the same program on the same input should always produce the same output. As large language models move into real-world deployment, this ideal has been imported wholesale into inference stacks. Recent work from the Thinking Machines Lab has presented a detailed analysis of nondeterminism in LLM inference, showing how batch-invariant kernels and deterministic attention can enforce bitwise-identical outputs, positioning deterministic inference as a prerequisite for reproducibility and enterprise reliability. In this paper, we take the opposite stance. We argue that, for LLMs, deterministic inference kills. It kills the ability to model uncertainty, suppresses emergent abilities, collapses reasoning into a single brittle path, and weakens safety alignment by hiding tail risks. LLMs implement conditional distributions over outputs, not fixed functions. Collapsing these distributions to a single canonical completion may appear reassuring, but it systematically conceals properties central to artificial cognition. We instead advocate Stochastic CHAOS, treating distributional variability as a signal to be measured and controlled. Empirically, we show that deterministic inference is systematically misleading. Single-sample deterministic evaluation underestimates both capability and fragility, masking failure probability under paraphrases and noise. Phase-like transitions associated with emergent abilities disappear under greedy decoding. Multi-path reasoning degrades when forced onto deterministic backbones, reducing accuracy and diagnostic insight. Finally, deterministic evaluation underestimates safety risk by hiding rare but dangerous behaviors that appear only under multi-sample evaluation.
Abstract (translated)
URL
https://arxiv.org/abs/2601.07239
PDF
-
MultiFlow: Vision-Driven Multimodal fMRI Encoding via Stimulus-Informed Source Flow Matching
2026-10-11 04:33:44Bui Van Hung, Hoang-Son Vo, Mai Duc Minh Huy, Dang Thanh Vu, Soo-Hyung KimAbstract
Current fMRI encoding models frame prediction as deterministic regression, yet identical stimuli evoke a distribution of neural responses (arising from physiological noise, hemodynamic variability, and cognitive fluctuations) that a single point estimate cannot capture. This work introduces MultiFlow, a generative fMRI encoding framework based on Conditional Flow Matching (CFM) and a formulation termed Stimulus-Informed Source Flow Matching (SISFM). SISFM generalises CFM by conditioning the source distribution on the stimulus: a learnable temporal convolution shapes the source mean with an HRF-length receptive field, while a variance head shrinks its scale, shortening the expected transport path. SISFM is instantiated with a Diffusion Transformer (DiT-X) decoder parameterising the velocity field, with sub-layer Attention Residuals, heterogeneous fusion of ten frozen vision, audio, language, and audiovisual backbones, and per-subject Yeo-network heads. On the Algonauts 2025 dataset, a single MultiFlow checkpoint recovers most of the benefit of large deterministic ensembles: it trails the 1000-model TRIBE ensemble in-distribution by 0.003 PCC (0.3163 vs. 0.3195), and the strongest in- and out-of-distribution ensembles (VIBE-20, MedARC-50) by 0.010 and 0.009 PCC, using a single trained model rather than an ensemble of 20-1000. The same checkpoint also provides stimulus-conditioned uncertainty estimates and sample-based diversity.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15339
PDF
-
RECAP: Calibrating Low-Bit Reasoning Models on the History They Read
2026-10-11 01:10:33Chi Ruan, Yujie Chen, Zixiang Zhao, Jingcai Guo, Haotong QinAbstract
Low-bit quantization reduces the cost of serving reasoning models but can substantially degrade long chains of thought. Each step reads cached keys and values spanning thousands of tokens, allowing quantization errors to accumulate along the trajectory. Standard calibration operates on short segments that omit the earlier context needed by later reasoning steps. Even when this context is retained, cache reconstruction error alone does not capture how quantization errors affect subsequent attention. We present RECAP, a post-training quantization method that calibrates reasoning models on the history they read. Dynamic-prefix calibration separates long historical context from short-block reconstruction. It regenerates prefix keys and values with the current quantizer and rotates the active block to supervise every position along the trajectory. Prefix read geometry further supervises historical cache errors according to their effects on attention outputs, weighting key and value errors asymmetrically. With both components applied only during calibration, RECAP outperforms the baseline across multiple reasoning models and quantization settings. Under W4A4KV4, RECAP improves accuracy over the baseline by 13.3 percentage points on MATH500 with DeepSeek-R1-Distill-Qwen-1.5B and 15.8 points on AIME-120 with Qwen3-4B. These results highlight the importance of preserving historical dependencies during calibration for low-bit reasoning.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15259
PDF
-
Understanding Unified VLMs: From Attention Probing to Guidance Tuning for Multimodal Generation
2026-10-10 23:40:27Zhengyang Luo, Xiaoyu Qiu, Zhiheng Tan, Yuchen Liang, Yingbin Liang, Liyue ShenAbstract
Unified vision-language models (VLMs) perform both text-to-image (T2I) and image-to-text (I2T) generation with one shared Transformer backbone, yet how the backbone processes the condition in either direction remains poorly understood. We study this question from the perspective of conditional guidance: we probe every attention head to show the activation difference between two branches of classifier-free guidance (CFG) with different conditional guidance. On two unified VLMs of autoregressive and masked-diffusion paradigms, the probing reveals a subset of heads that are more sensitive to the change of condition while others are agnostic to it. This characteristic pattern holds in both T2I and I2T generation and is largely stable across data samples, demonstrating the internal mechanism of unified VLMs. Motivated by this, we propose a new inference-time intervention approach for unified multimodal generation, named \textit{HeadTuned-CFG}, which uses the probed condition-change awareness to select and tune the conditional guidance at the attention-head level. Experiments show consistent benefits in both multimodal generation directions of unified VLMs without additional training, and in particular substantial improvements over strong baselines in T2I generation. These results further support the key finding of our probing analysis: the observed pattern of functional specialization is an intrinsic property of unified VLMs.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15227
PDF
