Attention
-
What LLMs Do in the Shadows: Reading Context Reliance from the Unembedding Dark Space
2026-10-16 09:21:49Alexandra Kuleshova, Andrey Volodichev, Daria Kotova, Alexey ZaytsevAbstract
Detecting faithfulness hallucinations in large language models (LLMs) is crucial for their safe deployment in high-stakes domains. Recent approaches address this by contrasting contextual against parametric knowledge utilization, yet measuring context reliance typically demands noticeable overhead, such as materializing attention maps, running additional forward passes, or training auxiliary probes. In this work, we show that context reliance can be estimated by projecting hidden states onto the dark space - directions that compose weakly with the unembedding matrix. Beyond its recognized function as an initial-token attention sink, we show that this subspace actively mediates context processing, and that its share of hidden-state norm - termed the Dark Matter Ratio (DMR) - provides a direct score of context utilization. Pairing DMR with a parametric knowledge score yields DarkLens, a lightweight faithfulness detector operating strictly within a single forward pass. Evaluated across three QA and summarization benchmarks on six open-weight LLMs (4B-13B parameters), DarkLens matches or outperforms a wide range of baselines without requiring extra forward passes or probe training.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23182
PDF
-
RealtimeWAM: One-Step Asynchronous World Action Models
2026-10-16 09:12:16Chengtao Lv, Jinyang Du, Shuyi Feng, Yang Yong, Shiqiao Gu, Shunzi Yang, Ruihao Gong, Shen Ren, Tianwei Zhang, Wenya WangAbstract
World Action Models (WAMs) incorporate visual representations from video generation backbones to guide action prediction. Recent efficient WAMs adopt Mixture-of-Transformers (MoT) architectures and compute video representations once for reuse by the action expert. However, intra-expert iteration (\ie, multi-step action denoising) and inter-expert waiting (\ie, sequential execution of the video and action experts) still limit inference efficiency. To this end, we present RealtimeWAM, an extremely efficient WAM variant with one-step action generation and asynchronous inference, addressing these two bottlenecks. To reduce intra-expert iteration, we propose Teacher-Anchored Consistency Distillation (TACD) to address a local-global error gap: low local consistency error alone does not guarantee accurate final actions. TACD supplements local consistency with explicit supervision from the frozen teacher's multi-step rollout endpoint, enabling accurate one-step action generation. Additionally, we propose Cross-Expert Wavefront Pipelining (CEWP) to eliminate unnecessary expert-level waiting. It overlaps the two experts through block-wise sharing of the video KV cache, synchronizing only immediately before the corresponding action attention consumes it. Extensive experiments across diverse benchmarks (\eg, LIBERO, LIBERO-Plus and RoboTwin) and model variants (\eg, Fast-WAM and Faster-WAM) demonstrate the superiority of RealtimeWAM. Notably, RealtimeWAM maintains near-lossless performance (\ie, $<1\%$ drop) across these benchmarks while delivering significant end-to-end speedup (\eg, $\sim25\times$ on H100). Our code and checkpoints are available via this \href{this https URL}{link}.
Abstract (translated)
URL
https://arxiv.org/abs/2610.06617
PDF
-
GRASP: Global-Local Representation Alignment and Spatial-Physical Grounding for Bimanual Robotic Manipulation
2026-10-16 08:29:19Zeyuan Ren, Guanxun Zhao, Qi Li, Hao Wang, Wenlong Liao, Keyi Luo, Xijing Zhang, Zhiyuan Zhang, Jingye Cui, Qibin Gu, Zhichen Wang, Yucheng Hu, Liqiao Xia, Nan Bu, Zhengde Ma, Rankun Li, Xue Yang, Wende XieAbstract
Vision-Language-Action (VLA) models that directly regress joint-space actions avoid the deployment-time cost of inverse kinematics. However, their training objectives remain misaligned with task success, as joint-space losses ignore the unequal physical consequences of joint errors and visual conditioning lacks explicit focus on hand-object interaction regions. To address these issues, we propose a training-phase prior injection framework for fine-grained bimanual manipulation. First, a physics-grounded flow-matching loss maps the implicit clean-action estimate at every noise level through differentiable forward kinematics (FK) and supervises the resulting Tool Center Point (TCP) pose and inter-arm relative pose, with a bounded noise-level weighting that compensates gradient decay. Second, we introduce a global prediction query group distilled from a frozen DINOv3 teacher through future-frame feature prediction, together with a local interaction query supervised by Grounding DINO at the attention level. Both teachers are removed at inference, and the deployment pipeline is that of a standard joint-space flow policy, adding only the learned query pathway. Simulation and real-robot experiments on long-horizon bimanual tasks show that the two priors are complementary across different deployment settings. The full model excels in in-distribution deployment, while the distilled query branch is preferable under strong visual randomization.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23123
PDF
-
PromptSep-Speaker: Generative Target Speaker Extraction via Multimodal Prompting and Semantic Alignment
2026-10-16 07:41:44Yutong Wen, Ke Chen, Prem Seetharaman, Haonan Chen, Oriol Nieto, Paris Smaragdis, Justin Salamon, Zeyu JinAbstract
Target speaker extraction (TSE) isolates a desired speaker from a multi-speaker mixture using cues that identify the target. Prior work has explored acoustic and textual cues, such as enrollment audio and different types of speaker descriptors, but typically relies on a single cue family. Yet no single cue is both broadly available and reliably discriminative. Moreover, most TSE systems adopt mask-prediction models, which are prone to audible artifacts and generalize poorly across acoustic conditions. To address these limitations, we propose PromptSep-Speaker, a generative TSE model based on latent diffusion that unifies four interchangeable cues: speaker temporal order, transcript, enrollment audio, and speaker caption. The generative formulation delivers higher separation quality and multi-modal prompting improves applicability across scenarios. To fully exploit each cue, we further introduce a frozen WavLM representation with an auxiliary alignment loss that supervises query-mixture cross-attention without extra labeling. Across four evaluation sets and a listening study, PromptSep-Speaker achieves state-of-the-art performance over TSE and universal source separation baselines, improving separation accuracy and quality across cue modalities.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23066
PDF
-
BRICE: Brain-inspired Recurrent Illusory Contour Encoder for Point Cloud Completion
2026-10-16 06:25:09Yunlei Wang, Dawei Li, Xue-song TangAbstract
Mainstream point cloud completion methods are unanimously based on deep learning architectures. On one hand, these methods tend to repetitively stacking attention modules to enhance network performance, which drifts away from the neat design of unified and interpret-able networks; on the other hand, nearly all existing architectures rely solely on feed-forward connections to propagate information, neglecting feedback from deeper layers. The recurrent neural architecture, formed by feedforward-feedback loops, have shown its ability to generate Illusory Contours (ICs) in the V1 brain region of mammals; and ICs are essentially a completion mechanism of the brain for visual information. Motivated by this, we design a Brain-inspired Recurrent Illusory Contour Encoder (BRICE) to "imagine" what is missing in the incomplete 3D point clouds. Building upon the foundation model of Predictive Coding (PC), the plug-and-play nature enables BRICE to be seamlessly integrated into mainstream point cloud completion baselines. Experimental results show that BRICE can enhance various completion baselines on different 3D datasets such as MVP2K, PCN, and PlantPCom. Our code: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23001
PDF
-
Prism Attention: Proposal-Refined Index Sharing Mechanism for Efficient LLMs Inference
2026-10-16 06:11:20Zhenxu Tian, Kebin Liu, Zhengwu Yang, Yi Su, Qingqing Dang, Yanlin Sha, Kaipeng Deng, Dianhai Yu, Yanjun Ma, Juntao Li, Min ZhangAbstract
Long-context decoding in large language models (LLMs) is increasingly bottlenecked by the memory and computation required to attend over an ever-growing Key-Value (KV) cache at each autoregressive step. Although token-level sparse attention reduces this burden by restricting computation to a small subset of relevant past tokens, its practical gains are often offset by repeatedly recomputing layer-wise sparse indices as the context grows. A common remedy is to share sparse indices across layers by computing them only at designated anchor layers, but such static reuse ignores the layer-wise evolution of attention patterns, causing the reused indices to progressively deviate from the layer-specific optima and incur substantial attention loss. By revisiting cross-layer sparsity from the perspective of attention dynamics, we find that inter-layer salient token shift is highly localized: although the exact top-$\kappa$ indices differ across layers, the core high-attention tokens remain stable, with changes concentrated near the expanded boundary of previously selected regions. Motivated by this, we propose Prism Attention, a proposal-refined sparse attention framework. It replaces rigid cross-layer index reuse with a coarse-to-fine strategy, preserving layer-wise adaptability while retaining the efficiency benefits of cross-layer sharing. Extensive experiments show that Prism Attention achieves state-of-the-art performance among mainstream sparse attention methods on both reasoning and long-context benchmarks, while attaining a $3.4\times$ end-to-end decoding speedup over FlashAttention at 128K context length.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22985
PDF
-
Sink-Sparse Hybrid Attention for Efficient Inference in T2I/T2V Diffusion Transformers
2026-10-16 05:50:56Bingxin Xing, Yu Zhang, Jiatong Ding, Lei Lu, Ping WangAbstract
Diffusion Transformers (DiTs) have become a common backbone for high-quality text-to-image and text-to-video generation, but repeated dense self-attention remains a major inference bottleneck as token sequences grow and the same operator is invoked across many denoising steps. Our study suggests that dense attention in pretrained DiTs contains reusable layer/head-specific high-mass structure that can be identified from a small offline calibration set and reused across future generations. This leads to Sink-Sparse Hybrid Attention, a train-free attention replacement that reuses static sink masks and supplements them with a lightweight pooled-KV branch, while leaving model weights and sampling schedules unchanged. On Wan2.1-T2V-1.3B, the method attains 1.54$\times$ end-to-end speedup while maintaining generation quality. On Z-Image-Turbo, it attains 1.57$\times$ end-to-end speedup with competitive GenEval2 performance. These results indicate that reusable importance-aware sparse support can provide an effective train-free acceleration path for both T2V and T2I DiTs.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22975
PDF
-
MambaAct: Efficient Mamba Action Experts for Flow-Matching Vision-Language-Action Models
2026-10-16 05:20:45Khuyen Pham, Tran Nguyen Le, Edoardo Fazzari, Cesare StefaniniAbstract
Vision-language-action (VLA) models have shown strong promise for generalist robotic manipulation, but their inference cost remains a major bottleneck for real-time control. This limitation is especially pronounced in flow-matching VLAs such as $\pi_{0.5}$, where the action expert is evaluated repeatedly during iterative denoising. We introduce MambaAct, an efficient Mamba action expert for flow-matching VLA policies. MambaAct retains the pretrained visual-language prefix encoder of $\pi_{0.5}$, while replacing its transformer-based action expert with a compact Mamba-based action expert. Within this expert, Mamba serves as the action mixer that models dependencies across the noisy continuous-action sequence. To preserve visual-language conditioning without repeated token-level cross-attention, MambaAct uses a lightweight Context Compressor to map Vision-language model (VLM) prefix tokens into a compact conditioning signal that gates and modulates each Mamba action block. MambaAct uses a $10\times$ smaller action expert and reduces flow-matching denoising latency by $64.8\%$ on LIBERO and $58.8\%$ in real-world deployment, while improving mean LIBERO success by $1.08$ percentage points across three independent fine-tuning seeds and maintaining gains on LIBERO-Plus and real-world benchmarks. Our code and model weights are publicly available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22951
PDF
-
Many Biases, One Circuit: Post-Training Tunes a Pretrained Cue-Routing Mechanism
2026-10-16 05:00:12Prakhar Gupta, Terry Jingchen Zhang, Zhijing JinAbstract
Language models often abandon a correct answer when the prompt contains a cue, a hint toward a wrong option such as "I think the answer is B." Sycophancy and related biases show this failure. They are often studied separately and blamed on alignment. We test this view causally, tracing the attention heads behind these biases through the stages of two post-training pipelines. In multiple-choice tests on most models, four cue types rely largely on the same few heads, a shared circuit already present in pretrained models. Most base models follow cues in plain text. Fine-tuning a base model on about 200 generic assistant examples without cues activates these heads in chat format, so cue-following can be a side effect of assistant tuning. The final reinforcement learning (RL) stage strengthens the effect of cues in almost every comparison. In Tulu-3-8B and OLMo-2-7B, we swap the attention patterns of six middle layers between the RL checkpoint and its predecessor. For suggested-answer cues, this shifts each checkpoint's cue effect most of the way to the other's level. Post-training thus seems to change when and how strongly models use this circuit. Mitigation for these multiple-choice biases could target this circuit.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22937
PDF
-
Sep-Latent: Separator Tokens as Natural Compression Anchors for Implicit Reasoning
2026-10-16 04:39:25Yinlong Xu, Jianfeng Huang, Zhengdong Zhou, Mingjie Lin, Fuyu Lv, Dan Ou, Hongxia Xu, Jian WuAbstract
Large Language Models (LLMs) have demonstrated remarkable success in complex reasoning through explicit Chain-of-Thought (CoT), but the associated computational overhead and token verbosity remain significant barriers to efficiency. While implicit reasoning in continuous latent spaces offers a promising alternative, existing methods like Coconut and CODI often rely on fixed-length latent sequences and suffer from representational collapse. In this paper, we propose Sep-Latent, an adaptive implicit reasoning framework that utilizes separator tokens as structural semantic anchors. By introducing an Attention-Decay mechanism, we force the model to aggregate and compress reasoning logic into these anchors during training. We then align explicit and implicit pathways through multi-level distillation across all layers to ensure logical consistency. Crucially, Sep-Latent enables dynamic reasoning depth by adaptively terminating the latent thinking phase based on query complexity. Experimental results on the GSM8k-Aug benchmark show that Sep-Latent achieves an accuracy of 56.8\% with an average of only 5.4 tokens, significantly outperforming prior fixed-length implicit models in efficiency while narrowing the performance gap with explicit CoT. Further analysis via PCA and symbolic variant testing confirms that our method ensures semantic grounding and exhibits superior robustness to perturbations.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22924
PDF
-
FederatedBytes: Tokenizer-Free, Byte-Level Small Language Models for Efficient Multimodal Federated Learning
2026-10-16 04:14:35Ali Sarabadani, Massimo Piccardi, Mehran Abolhasan, Farzad Tofigh, Ian Zhou, Justin LipmanAbstract
Federated Learning (FL) is a powerful machine learning framework that allows building a global model from the contributions of multiple, distributed clients, while ensuring the privacy of their local data. However, its heavy computational and communication overheads challenge its use for contemporary AI workloads such as natural language processing (NLP) and multimodal analysis. For this reason, this paper presents a novel small language model that leverages byte-level text and multimodal data processing, enhanced by a linear attention mechanism, to enable manageable Federated Learning of NLP and multimodal analysis tasks. The proposed model, nicknamed FederatedBytes, manages to process text and any other data types, including images, by treating its input as a mere sequence of bytes, while its lean linear attention mechanism allows abating the overall memory and computational footprint. To prove the effectiveness of the proposed model we have carried out an extensive performance evaluation on sentiment analysis tasks in multilingual and multimodal settings over the Multilingual Amazon Reviews and ScienceQA datasets. The experimental results show that FederatedBytes has been able to outperform comparable models such as ByT5 and LSTM, achieving higher accuracy while significantly reducing memory and communication overheads.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22906
PDF
-
ToPA: Block-wise Toeplitz Adaptation for Expressive and Efficient Fine-Tuning
2026-10-16 03:38:18Sicong Li, Qianqian Xu, Zhiyong Yang, Zitai Wang, Longtao Huang, Xiaochun Cao, Qingming HuangAbstract
Parameter-Efficient Fine-Tuning (PEFT) has emerged as a prevalent strategy for adapting large pre-trained models. Among PEFT techniques, reparameterization-based methods, particularly those that employ low-rank or sparsity structures, have attracted significant attention. However, these approaches often underperform full fine-tuning. Specifically, low-rank methods constrain the update rank, which is misaligned with the inherently high-rank nature of full fine-tuning. In contrast, sparsity-based techniques limit the flexibility of the parameter space and compromise connectivity across weights. To overcome these limitations, we explore the use of Toeplitz matrices, whose entries are constant along each diagonal, therefore providing a compact parameterization without imposing explicit rank or sparsity constraints. A straightforward approach is to update weights via a product of Toeplitz matrices. However, similar to RNNs, long multiplicative chains often lead to gradient instability. To address this issue, we propose \textbf{ToPA}, a more refined version of this idea that replaces standard Toeplitz matrices with block-wise Toeplitz, greatly increasing each factor's capacity and shortening the chain, thereby stabilizing training and improving expressiveness. We theoretically show that products of block-wise Toeplitz matrices can approximate arbitrary matrices under shorter chain length, justifying the structural design of ToPA. Furthermore, we demonstrate that ToPA offers greater expressivity than existing low-rank and sparse parameterizations. Empirical evaluations across 21 NLP and CV datasets, spanning 5 model architectures, consistently validate the effectiveness of ToPA.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22883
PDF
-
Measurement-Adaptive Light-Cone Transform for Robust Non-Line-of-Sight Imaging
2026-10-16 02:47:58Lianfang Wang, Xueying Liu, Junnan Dong, Qinghua Zhang, Jun Liu, Yong Wang, Yuping DuanAbstract
Practical non-line-of-sight (NLOS) measurements are affected by noise and deviations from idealized light transport, which can degrade reconstruction with conventional light-cone transform (LCT) methods. We propose the Measurement-Adaptive Light-Cone Transform (MALCT), a physics-guided framework that retains the efficient Fourier-domain inversion structure of LCT. {Within a unified variational formulation, MALCT combines the analytical response with a shared learned response through measurement-conditioned weighting and predicts a measurement-conditioned, frequency-dependent regularization map.} An attention-enhanced spatio-temporal feature extraction module captures long-range temporal dependencies and local spatial correlations to support reconstruction under low-signal conditions. Experiments on synthetic data and real measurements from different NLOS systems demonstrate improved reconstruction quality. Controlled evaluations further assess robustness to temporal-response broadening and reduced acquisition time. MALCT also preserves scene structures with spatial sampling grids down to $32\times32$, supporting its potential for NLOS imaging under limited acquisition budgets.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22847
PDF
-
Causal Edit Serialization
Abstract
Standard autoregressive language models generate text sequentially, but revision requires referring back to tokens in the input context or those already generated. Causal Edit Serialization (CES) is introduced as a framework for expressing such revisions within the same causal token stream. Special pointer tokens are paired with predicted source addresses, allowing insertion, deletion, replacement, copying and movement to be localized explicitly. A lightweight localization head and dual-index rotary position encoding support both address prediction and conditioning on previously emitted pointers, while retaining a standard causal attention mask. A pretraining procedure combines ordinary language modeling with serialized repairs derived from random corruptions, model-generated edits and repeated text. Across model scales and editing baselines, CES models demonstrate complete applied edits with modestly higher clean-text bits per byte than clean autoregressive models. Supervised and reinforcement-learning post-training provide a grammatical-error-correction proof of concept, producing more beneficial than harmful changes under reference-blind model judging. Correct repair programs also improve subsequent text prediction while the original context and its cached states are retained. CES provides a common interface for generation and localized revision, with potential extensions to iterative refinement and additional pointer-addressed transformations.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22820
PDF
-
Causal Latent Workspaces for Looped Transformers
Abstract
Reusing a Transformer block across depth increases computation without replicating its parameters, but successive passes have no dedicated state through which positions can exchange compact summaries of their causal prefix. We introduce the causal latent workspace Transformer: an anchored, four-round shared-depth architecture in which each position writes gated vectors into prefix-specific latent slots and retrieves a content-addressed summary before the next shared block application. This path complements causal self-attention while the original input embedding remains accessible at every round. Across three byte-level corpora, three seeds, and two matched parameter budgets, the proposed model achieves the best mean held-out perplexity in five of six settings against five Transformer controls. At 101.71M parameters it leads on WikiText-103 raw, enwik8, and TinyStories, reaching perplexities 2.7820, 2.7843, and 1.6823; these improve upon the strongest non-workspace control on each corpus by 1.06%, 0.73%, and 1.01%. All nine seed-labelled comparisons at this budget favor the workspace. At 0.918M parameters it leads on WikiText-103 and TinyStories, whereas a conditioned loop is stronger on enwik8. Architectural ablations show that adding the workspace improves upon the anchored loop in all six settings and support refreshing its causal summaries at each round. Together, these results identify a causal, content-addressable prefix workspace as an effective structural augmentation of looped Transformers.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22796
PDF
-
Reading the News Without Hindsight: NoLBERT Topics and the Cross-Section of Stock Returns
2026-10-15 23:27:49Kevin Foley, Jonathan Hartadi, Shivesh Prakash, Charishma Takkallapalli, Swapnil VatsalAbstract
Language models that read financial news are usually trained on text from the same years used to test them, so a backtest can reward hindsight. We build news topics from NoLBERT, an encoder pretrained only on text published between 1976 and 1995, adapted into a sentence encoder by contrastive fine-tuning on passages of the same article, with no labels or returns. We compare it with LDA, MiniLM and FinBERT readers on 1.26 million articles from 2012 to 2024, convert each reader's topics into attention shocks and stock exposures, and forecast the returns of the 1,000 largest US stocks with eight models at three rebalancing frequencies. Models are fitted on 2013-2019, choices are made on 2020-2021, and 2022-2024 is held out. The experiment with the best validation score loses money at every frequency. Measuring attention shocks against the past 63 trading days instead of 5 raises the test rank IC of every reader. NoLBERT gives the most stable topics and forecasts about as well as LDA. An equal-weight combination of all readers and models earns a monthly net Sharpe ratio of 1.12, significant against zero before adjusting for the search but not against a price-only benchmark.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22745
PDF
-
LEMON: Equivariant Learning over Attention Maps for Zero-Shot Misbehavior Monitoring on New LLMs
2026-10-15 22:53:45Guy Bar-Shalom, Mor Geva, Fabrizio Frasca, Haggai MaronAbstract
Large language models exhibit misbehaviors that hinder deployment in high-stakes settings. One approach to monitor such failures is to use an external LLM as a monitor, but this is prohibitively expensive. A well-known and efficient class of monitors can detect such failures from internal representations (such as activation probes), but are tied to the model they were trained on, requiring costly re-annotation as new LLMs appear. A natural way to avoid this repeated annotation cost is to develop internal-representation based monitors that transfer zero-shot across LLMs. However, such transfer is challenging because internal representations are not naturally aligned across models; in particular, attention maps, known to contain signals of misbehavior, may differ in head function, ordering, and model depth. In this work, we characterize those cross-model attention mismatches as algebraic symmetries and use geometric deep learning principles to build LEMON (Language-model Equivariant MOnitoring Network) an online graph neural network-based monitor, whose predictions are invariant under these symmetries. LEMON processes attention maps token-by-token, as they are produced and outputs a score after every token (or every window of $W$ tokens) at negligible cost. Across 16 LLMs and six datasets, in the zero-shot cross-model setting, LEMON outperforms all baselines on both hallucination and jailbreak monitoring tasks, improving over the strongest by up to 14 ROC-AUC points.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22723
PDF
-
Risk and Refusal: Mechanisms of LLM Responses to Suicidal Ideation
2026-10-15 22:34:45Drishti Goel, Soorya Ram Shimgekar, Jeongah Lee, Violeta J. Rodriguez, Daniel S. Brown, Ravi Karkar, Koustuv SahaAbstract
Refusal is a common safety mechanism in large language models (LLMs), yet in contexts of suicidal ideation (SI), refusing to engage can itself constitute a safety failure. We evaluate how eight instruction-tuned LLMs respond to SI and self-harm disclosures across three datasets, varying whether models are informed that a disclosure expresses SI and whether they receive a clinician-developed response protocol, and complement this with mechanistic analyses in six open-weight models. Our results reveal a nuanced relationship between refusal and response-level risk. At baseline, elevated risk in model responses is driven primarily by inattention to the disclosure, alongside refusal to engage. Mechanistic analyses suggest that models may refuse SI disclosures through the same direction that mediates their refusals of harmful requests. However, informing models of the SI context, or suppressing refusal directly, increases engagement without reliably lowering response risk, and may raise it. In contrast, the clinician-guided protocol consistently reduces response-level risk while also reducing refusal. Together, these analyses and ablations motivate the need for models that refuse only when warranted and engage in interactionally safe ways when users disclose SI.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22715
PDF
-
First Passage Attention
Abstract
Augmenting softmax attention with forget gates typically relies on summing decay contributions along the key--query path. Because such accumulation is monotonic, these mechanisms cannot selectively erase intermediate context while preserving older, structurally relevant tokens. We propose First Passage Attention (FPA), which shifts from path summation to path extrema by lowering a key's logit proportionally to the deepest dip of a learned scalar field below that key's level over the intervening span. FPA encompasses the Forgetting Transformer (FoX) as a special case, and when the field and levels follow a shared learned walk, the alive keys with unpenalised logits naturally form a stack. We show that a constant-dimension FPA head realises any subset of keys left alive by subsequent dips, whereas softmax heads with positional biases or token-level forget gates require a head dimension linear in the number of keys. In synthetic experiments, one-layer FPA heads converge to these target sets while comparable softmax and FoX baselines fail on every seed. Scaling to a 110M-parameter language model, FPA matches FoX in perplexity and uniquely achieves perfect passkey retrieval at $8\times$ the training length, even with its key--value cache compressed to about $25\%$.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22714
PDF
-
Brain-Aligned Multimodal LLMs for Social and Emotional Understanding of Cartoon Movies
2026-10-15 22:22:03Yiran Huang, Zeyong Zhang, Bharat Biswal, Xin Di, Mengjia XuAbstract
Multimodal large language models (MLLMs) offer scalable annotations of social and affective content, but their temporal agreement with human judgments and neural predictive value require further evaluation. We compared diffusion-based and autoregressive MLLMs with continuous human Theory of Mind (ToM) and affective valence ratings across three short cartoon movies. Behavioral analyses assessed temporally aligned Pearson correlations, rating distributions, and the effects of preceding scene descriptions. Ratings from each source were separately used in subject-specific regression models to predict regional fMRI responses. Model ratings captured aspects of human temporal variation despite systematic distributional differences, with generally stronger alignment for valence than ToM. Preceding context improved alignment in some conditions, depending on the model and movie. In \emph{The Present}, more regions showed significant encoding for MLLM ratings than for human ratings, whereas effects were more limited in the other movies. Spatial overlap across rating sources occurred mainly among MLLMs; overlap between human and model ratings was limited and involved primarily visual, dorsal-attention, and posterior default-mode regions. These findings support MLLMs as scalable tools for generating brain-predictive annotations of social and emotional content, while highlighting systematic rating differences and model- and stimulus-dependent alignment with human judgments.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22706
PDF
