Sparse
-
Highlight-Then-Summarize: Learning to Compress Evidence for Long-Context Understanding
2026-09-25 15:16:35Zhaoyuan Xia (Peking University, Baidu Inc), Qinghongbing Xie (Tsinghua University), Yung Xiang Hue (Tsinghua University), Jianguang Jiang (Baidu Inc), Gaofeng Lu (Baidu Inc), Zhenyu Jiao (Baidu Inc), Xing Yuan (Baidu Inc), Dai Dai (Baidu Inc), Tong Mo (Peking University), Long Zeng (Tsinghua University)Abstract
Long-context understanding requires large language models (LLMs) to reason over lengthy documents, conversations, and code, yet task-relevant evidence is often sparse and scattered amid substantial irrelevant and redundant content. We propose Highlight-Then-Summarize (H2S), a compress-then-reason paradigm that first identifies source-grounded, question-relevant evidence and then integrates it into a compact, question-conditioned summary before producing the final answer. To train this behavior, we construct H2S-Dataset, comprising 6,647 examples from 11 benchmark families with an average context length of 43.9K tokens, and introduce H2S-RL, which provides process-level rewards for evidence selection and summary construction in addition to final-answer correctness. We evaluate on H2S-Bench, a seven-task long-context suite. Under a shared 128K input and 4K output budget, H2S-14B achieves an average score of 32.60, outperforming Qwen3.8-27B by 10.17 points and obtaining the strongest overall result among the evaluated open-source models. H2S-14B also achieves the highest Evidence-Summary Quality score and retains 97.1% of its 16K-budget performance with only a 4K output budget. These results show that explicitly selecting and integrating evidence improves long-context reasoning while enabling more compact generation.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31382
PDF
-
RECAST: From Log Replay to Closed-Loop Driving Simulation with View-Complete Actors
2026-09-25 15:12:05Zijun Zhao, Liewen Liao, Kang Shen, Songan Zhang, Ming YangAbstract
Closed-loop driving simulation requires rendered observations to remain reliable as the ego vehicle and surrounding actors move beyond their recorded trajectories, exposing views absent from the source log. Existing data-driven simulators reconstruct dynamic actors from sparse observations, which can result in rendering artifacts under these viewpoint changes. We introduce RECAST (REconstructing Controllable Actors for Simulation and Testing), a 3D Gaussian Splatting framework that generates a view-complete actor from a single segmented vehicle observation in a driving log and registers the generated actor in the reconstructed scene. RECAST supports planner-in-the-loop rendering under controlled ego-actor interactions. To adapt an image-to-3D prior to real vehicles, we further introduce RECAR, a dataset of approximately 20K real vehicles with 600K background-free RGBA images spanning diverse vehicle colors and types. We use two-stage adaptation to improve vehicle generation from real driving-log observations. At the actor level, RECAST reduces $\mathrm{FD}_{\mathrm{incep}}$ from 9.788 to 7.992 relative to unadapted TRELLIS. At the scene level, under actor motion beyond logged trajectories, RECAST reduces $\mathrm{FD}_{\mathrm{incep}}$ from 129.35 to 112.10 and increases $\mathrm{CLIP}_{\mathrm{margin}}$ ($\times1000$) from 0.14 to 3.47 relative to Street Gaussians. We demonstrate planner-in-the-loop simulation with the image-conditioned planner GTRS-Dense. Compared with native Street Gaussians actors, RECAST increases the no-collision (NC) rate from 22.2% (12/54) to 63.0% (34/54) and the mean minimum predicted time-to-collision (TTC) from 0.798 s to 2.150 s. These experiments show that RECAST supports closed-loop planner evaluation under controlled ego-actor interactions beyond log replay. Visit our project page at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2609.31374
PDF
-
MoSAR: Mixture of Semantic Attention Regimes for Learning Adaptive and Approximable Attention Geometries
2026-09-25 13:38:28Michele Paolicelli, Alessandro Petruzzelli, Alessandro Franceso Maria Martina, Cataldo Musto, Giovanni SemeraroAbstract
The quadratic complexity of dense self-attention remains a central bottleneck for long-context language modeling. Many efficient alternatives address this cost by deciding in advance where attention should be sparse or local. We argue that attention approximation should instead be approached as a geometric problem, with the relevant interaction geometry learned from data: natural-language dependencies are input-dependent and difficult to prescribe in advance, so the model should learn where positional relevance can decay and where broader interactions must be preserved. We introduce Mixture of Semantic Attention Regimes (MoSAR), which learns such an adaptive, controlled-decay geometry over query--key interactions. Input-conditioned query and key routers, applied after positional encoding, select mixtures over short, medium, and global regimes, inducing a continuous distance-dependent attention field rather than a fixed sparsity pattern. This geometry is learned during training and can subsequently be discretized through top-1 routing. In controlled pre-training experiments with matched 500M-parameter models, MoSAR learns a substantially lower-reach attention geometry without degrading language-modeling quality, improving perplexity over dense RoPE at the training context length. Under length extrapolation, MoSAR achieves the best perplexity among all evaluated variants, including strong baselines such as ALiBi. Moreover, the learned geometry remains stable under deterministic top-1 discretization, suggesting that it is not only adaptive, but also amenable to low-cost approximation at inference time.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31261
PDF
-
AtomWorld-Mem: Memory-Restored World States for Long-Horizon Atomistic Evolution
2026-09-25 11:26:05Tian Luo, Ruge Zhang, Haozhi Han, Yifrng Chen, Yunquan Zhang, Yunxin Liu, Ting Cao, Kun LiAbstract
High-fidelity atomistic evolution over long timescales requires more than observing the current crystal configuration. Instantaneous atomistic snapshots are often incomplete: locally similar configurations can correspond to different hidden dynamical contexts, future event preferences, and waiting-time scales. We argue that this snapshot ambiguity makes long-horizon atomistic evolution fundamentally a memory-based world-state restoration problem. To address this, we introduce AtomWorld-Mem, a memory-restored atomistic world model that recovers the latent world state missing from instantaneous crystal snapshots. AtomWorld-Mem treats the evolving alloy as an AtomWorld: spatial encoders write multi-scale atomistic keyframes from dense local topology and sparse long-range defect context, while short-term event memory and long-term structural memory integrate these keyframes across time to restore a future-predictive evolutionary state. The restored state is used to prioritize legal vacancy-mediated events under single-event Kinetic Monte Carlo (KMC) constraints, while event legality, physical execution, and residence-time updates remain governed by the underlying simulator. Empirically, AtomWorld-Mem improves long-horizon atomistic progress under fixed microscopic event budgets while maintaining high-fidelity evolution across energetic, structural, and vacancy-transport observables. It further transfers zero-shot across diverse unseen alloy-temperature AtomWorlds, suggesting that the learned memory-restoration mechanism captures reusable principles of hidden-state inference rather than a system-specific local energy heuristic. These results position memory-restored world-state modeling as a promising route toward efficient, physically grounded, and transferable atomistic evolution.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31133
PDF
-
LocUS: Head Selection and Subspace Projection for Targeted Activation Steering
2026-09-25 11:13:24Irene Tallini, Lorenzo Basile, Valentino Maiorca, Francesco Locatello, Alberto CazzanigaAbstract
Activation steering is a powerful training-free paradigm for controlling large language models at inference time. However, standard approaches estimate a per-layer steering direction from contrastive data and apply it on the layer's entire representation space, which may couple the intervention to off-target properties present in the contrastive data and degrade unrelated capabilities. To mitigate this issue, we introduce LocUS (Localized Unembedding Steering), a method which grounds activation steering to the model's own output vocabulary subspace. By identifying a property-specific linear subspace within the unembedding matrix, LocUS enforces a geometric constraint that restricts the steering transformation to a specific subspace and at the same time localizes its application to a sparse subset of attention heads. Extensive evaluations across three model families on toxicity mitigation, sentiment redirection and sycophancy suppression show that LocUS matches or outperforms state-of-the-art baselines while intervening on under 6% of parameters and better preserving general capability.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31122
PDF
-
Neuralyzing the Trace: Selective Representation-Level Unlearning with Contrastive Sparse Autoencoders
2026-09-25 09:51:52Itai Zehavi, Fanny Jourdan, Ulrich AivodjiAbstract
Machine unlearning aims to remove targeted information while preserving a model's other abilities. In realistic settings, such as privacy requests under the EU GDPR, the target may be narrow, for example information associated with a single person. Behavioral forgetting alone may be insufficient, motivating interventions directly on internal representations. However, standard mechanistic-interpretability extractors are poorly selective for such targets. We identify an energy bias in reconstruction-based extraction, which favors dominant background structure over low-energy target-specific components. We introduce SCALPEL, a contrastive sparse autoencoder designed to learn more selective forget features. We show theoretically that contrastive training promotes target-selective features and that our selection score controls expected background knowledge perturbation. We validate SCALPEL experimentally on TOFU across Qwen, Llama, and Gemma, where it substantially improves over NMF and standard SAE interventions and is competitive with Gradient Difference and RMU, bridging mechanistic interpretability and fine-grained unlearning.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31056
PDF
-
TRACKGRAPH: Online Open-Vocabulary 3D Scene Graphs via Image-Space Tracking
2026-09-25 08:51:33Peder Borge Hellesylt, Albert Gassol Puigjaner, Kostas Alexis, Annette StahlAbstract
Open-vocabulary 3D maps enable robots to reason about previously unknown environments using natural language. However, existing systems typically segment every incoming image, associate detections with persistent 3D segments, and frequently perform costly Vision-Language (VL) inference. We present TRACKGRAPH, an online open-vocabulary system that maintains short-term 2D mask identity directly in the image stream before fusing segments into 3D. FastSAM masks and CLIP features are computed at sparse keyframes, while dense DINOv3 features are used to propagate masks at a high rate in between. The resulting tracked masks are fused into a class-agnostic 3D segment layer within a hierarchical scene graph, with 3D association handling tracking interruptions and long-term revisits. Compact multi-view CLIP embeddings enable open-vocabulary retrieval. Across Replica, ScanNet++, and HM3D, TRACKGRAPH achieves competitive open-vocabulary segmentation and retrieval against state-of-the-art mapping methods, including the highest synonym frequency on Replica (0.50). On the same NVIDIA A100, it is 1.7x faster and uses 3.3x less GPU memory than ViT-H OVI-MAP. Real-world quadruped deployments demonstrate onboard scene graph construction and object search at 7.5Hz, while recorded drone data is used to test the method under aerial viewpoints.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31005
PDF
-
FTB Graph: Determining and Validating First-token Broadcasters and Language-Identity Head Circuits in Multilingual Language Models
2026-09-25 08:05:03Arjun Pillai, Christian Hoang, Anjelo LarozaAbstract
Large language models operating in multilingual contexts must resolve target response languages early in generation, yet the causal circuitry governing first-token language identity decisions remains poorly mapped. We present an end-to-end structural circuit analysis across six model architectures spanning four families: GPT-2, BLOOM-560M, Pythia-1B/2.8B, and Qwen2.5-1.5B Base/Instruct. Using Edge Attribution Patching (EAP) with FP16 active clamping, followed by exact activation patching verification with a 2,000-candidate-edge search ceiling, we extract directed acyclic graphs driving first-token language broadcasting. Across the standalone models, we observe deep or mid-to-deep broadcasting hubs, though the evidence is strongest for Pythia-2.8B and BLOOM-560M because GPT-2 and Pythia-1B leave few out-of-graph heads for comparison, while both Qwen2.5-1.5B variants invert the necessity check. Scaling from Pythia-1B to 2.8B expands node participation while maintaining a similar verified edge budget, producing sparser topology. The Qwen2.5-1.5B base and instruct circuits retain 84.7% Jaccard similarity, including the Layer 27 hub, indicating that first-token routing is largely established during pretraining and preserved by instruction tuning. Finally, EAP scores correlate weakly with exact patching deltas across most models, showing that linear gradient approximations can diverge from causal interventions in FP16 and motivating exact-patching verification for reliable circuit discovery.
Abstract (translated)
URL
https://arxiv.org/abs/2609.30954
PDF
-
DAPEVO: Deep Adaptive Patch Frame-Event Visual Odometry
2026-09-25 07:59:28Luca Gandolfi, Simone Nascivera, Roberto Pellerito, Rong Zou, Chiara Plizzari, Davide ScaramuzzaAbstract
Visual odometry is essential for autonomous navigation in GPS-denied environments, yet RGB-based methods remain vulnerable to motion blur, challenging illumination, and dropped frames. Event cameras complement conventional cameras with high temporal resolution and dynamic range, but their asynchronous measurements complicate reliable correspondence estimation. We present DAPEVO, a learned visual odometry system that estimates image and event correspondences independently at shared patch locations and fuses their correlation evidence before motion refinement. Each tracked patch maintains image and event descriptors, and a learned scalar gate combines modality-specific correlation embeddings for each patch--frame edge before a shared recurrent refinement and bundle-adjustment update. DAPEVO also supports event-only observations, enabling continued tracking when RGB frames are sparse or unavailable, while modality-aware keyframe culling preserves scarce frame constraints. On UZH-FPV, when retaining only one in six RGB frames, DAPEVO's mean absolute trajectory error (ATE) increases by only 36%, from 1.00 to 1.36m, whereas the ATE of DPVO and RAMP-VO rises by factors of $3.7\times$ and $3.1\times$, respectively. On TartanEvent, DAPEVO similarly remains below 1m ATE at 3Hz RGB input, while DPVO and RAMP-VO exceed 9m. Under degraded RGB input on TartanEvent, DAPEVO achieves an ATE of 0.60m, compared with more than 4m for both DPVO and RAMP-VO, while also outperforming event-only DEVO at 0.87m.
Abstract (translated)
URL
https://arxiv.org/abs/2609.30947
PDF
-
Why Jailbreaks Succeed in Diffusion Language Models: An Energy Landscape Analysis
2026-09-25 05:35:36Thong Bach, Dung Nguyen, Thao Minh Le, Truyen TranAbstract
Existing attacks and defenses for diffusion-based large language models (dLLMs) target specific vulnerabilities but lack a shared framework explaining why attacks succeed. We propose one by interpreting safety alignment as shaping the denoising energy landscape: a well-aligned model routes harmful queries toward safe outputs through an energy barrier that separates the two regions. Current jailbreak attacks reduce to two strategies for circumventing this barrier: obscuring the query's safety disposition at initialisation, or intervening mid-trajectory to force the denoising path across the energy barrier. From this perspective and the result that masked diffusion models minimise kinetic energy during denoising, we derive three complementary, training-free detection signals: a step-0 ratio that reads the initial safety disposition from the logit distribution before generation begins, and two trajectory-velocity signals that track kinetic energy in complementary subspaces of the logit space. An attack must either reveal its intent at initialisation or expend kinetic energy to cross the barrier in at least one monitored subspace, so the three signals cover each other's blind spots in the energy budget by construction. Evaluation across three dense dLLMs (LLaDA-8B, LLaDA-1.5, Dream-7B) and a sparse mixture-of-experts dLLM (LLaDA-MoE-7B) confirms this complementarity. In stress tests of known attacks, every configuration that evades detection also fails to produce harmful content, suggesting that the detection and barrier-crossing thresholds are hard to separate.
Abstract (translated)
URL
https://arxiv.org/abs/2609.30841
PDF
-
Dialogue-Based Streaming Audio-Visual Target Speaker Extraction with Predictive Dialogue Information
2026-09-25 03:56:44Shuhan Zhang, Wenxuan Wu, Haizhou LiAbstract
In face-to-face, real-time communication, a talking agent must track the target speaker through natural pauses, turn-taking, and backchannels, often amid background cross-talk. Most target speaker extraction (TSE) studies, however, rely on simulated mixtures with full or sparse overlap and ignore the turn-taking of real conversations. We therefore introduce, to our knowledge, the first benchmark for online audio-visual TSE (AV-TSE), built from intact dyadic interactions with independent third-party interference. Observing that anticipating upcoming activity from semantic, acoustic, and facial cues benefits online AV-TSE, we propose an LLM-based target-speaker voice activity projection (TS-VAP) module. Unlike conventional VAP with separated speaker channels, it forecasts the future activity of the target and conversational partner directly from the overlapping mixture, drawing on the linguistic and conversational knowledge of a speech-LLM, and uses this prediction to guide a low-latency separator. We further combine this predictive context with historical and synchronous speaker context. Experiments show that TS-VAP consistently improves streaming extraction across multiple AV-TSE backbones, and that further combining historical, synchronous, and predictive context yields nearly 1 dB gain on real AV conversations. Project page: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.30774
PDF
-
EviDETR: Preserving Query-Relevant Temporal Evidence for Moment Retrieval and Highlight Detection
2026-09-25 02:53:54Haoran Sun, Yufan Li, Qichen Zhang, Haoran Zhao, Shuqi WangAbstract
Joint video moment retrieval and highlight detection requires identifying query-relevant temporal segments while estimating clip-level saliency, yet DETR-style pipelines do not explicitly preserve query-relevant evidence throughout encoding, decoding, and cross-task prediction. We propose EviDETR, an evidence-preserving framework with three components. Semantic-aware Feature Reweighting (SFR) enhances query-relevant clip representations through saliency estimation and cross-modal interaction. A Temporal Top-2 Mixture-of-Experts (TTop2MoE) decoder performs query-adaptive refinement via sparse expert routing. MR-to-HD (MR2HD) fusion transfers span-level retrieval evidence to clip-level highlight prediction through confidence-weighted multi-scale aggregation. Using CLIP+SlowFast features, EviDETR achieves 69.29 R1@0.5, 54.77 R1@0.7, and 48.41 Avg. mAP for moment retrieval on QVHighlights, together with 41.83 HD-mAP and 68.33 HIT@1. Strong results on TACoS and Charades-STA further demonstrate cross-dataset transferability.
Abstract (translated)
URL
https://arxiv.org/abs/2609.30724
PDF
-
QSV: Quat-Sphere-Vision for Coupled Quaternion Attention on Spherical Lattices
2026-09-24 22:13:57Nicholas Foley, Devin Marinelli, Donny Moore, Diego Enriquez, Amanda FernandezAbstract
In standard attention, three separately learned projections decide how strongly a token attends to each neighbor ($W_Q$, $W_K$) and how the attended features are transformed before aggregation ($W_V$). We study Quat-Sphere-Vision (QSV), a sparse spherical vision model that replaces this projection triple with a single learned unit quaternion per token: the relative quaternion $r_{ij} = q_i^{*} \otimes q_j$ supplies both the attention logit $\operatorname{Re}(r_{ij})$ and a sandwich-product feature transport $x \mapsto r_{ij} \otimes x \otimes r_{ij}^{*}$, with messages passed over sparse kNN graphs on concentric Fibonacci spheres. Ablations that change only the targeted component show the two roles to be asymmetric. Removing the transport reduces test accuracy by about four percentage points on CIFAR-10 and CIFAR-100 (single runs per CIFAR-100 variant), while replacing the learned attention weights with uniform averaging leaves it essentially unchanged. Parameter-matched controls then remove the geometry itself: standard attention on the same graph exceeds QSV (mean $87.3\%$ vs. $85.9\%$), and the same model on a flat 2D lattice reaches $91.1\%$, within $2.1$ points of a ResNet-20 trained under the same pipeline (single run). In the coupled kernel, nearly all of the learned pairwise computation resides in the transport channel.
Abstract (translated)
URL
https://arxiv.org/abs/2609.30592
PDF
-
T-RoPE: Time-Aware Rotary Position Embedding for Sequential Recommendation
2026-09-24 21:39:12Yang Liu, Noel Loo, Ali Khanafer, Shuying Sun, Akshay Soni, Zhong Wu, Linjun YangAbstract
Large-scale recommenders increasingly adopt the sequential generative recipe behind large language models, bringing the Transformer into recommendation along with design choices made for text, including Rotary Position Embedding (RoPE). In language models, RoPE encodes token indices for relative position reasoning, but in recommendation, an interaction index records only event order, saying nothing about elapsed time, behavioral cycles across scales, or calendar phase. We revisit this choice and propose T-RoPE, a time-aware RoPE for sequential generative recommendation that replaces index-only rotation with timestamp-based angles, learnable temporal coefficients, multiscale frequency banks, shifted query alignment, and non-stationary key rotation. We prove that standard RoPE, even on timestamps, remains time-translation invariant and cannot distinguish seasonal contexts, and that T-RoPE breaks this invariance while preserving the RoPE interface. Across five public benchmarks, T-RoPE achieves the best result on every metric on every dataset, improving over the strongest baseline by 78--130\% in HR@10 on the sparse PixelRec data and 8--12\% across metrics on Amazon Books. On an industrial-scale e-commerce dataset with more than 6B interactions, it improves every metric over the HSTU + Time RAB backbone by 13--82\%, with ablations attributing the largest gains to multiscale frequencies ($+56\%$ NDCG@50) and non-stationary keys ($+4\%$). An online A/B test in the Shop app yields positive lifts in conversion rate ($+0.33\%$) and order count ($+0.63\%$). We also provide forward and backward algorithms whose added cost is linear in sequence length and head dimension, keeping time-aware RoPE practical for large generative recommenders.
Abstract (translated)
URL
https://arxiv.org/abs/2609.30576
PDF
-
Spectral Feedback for Test-Time Alignment of Protein Diffusion Models
2026-09-24 18:49:51Shai Dickman, Mert Cemri, Landon Butler, Kannan RamchandranAbstract
Reward maximization alignment methods for discrete diffusion models have primarily focused on steering the reverse process, either by influencing token logits or by selecting favorable sequences at intermediate steps. These approaches largely treat inference as a unidirectional process, lacking mechanisms for revisiting undesirable token selections. We introduce Spectral Feedback, an algorithm that selects edit-positions in a feedback loop, allowing the model to iteratively correct its own generations. This approach leverages the mask structure of discrete diffusion models by re-masking and re-sampling tokens, analogous to image editing methods that reintroduce noisy latents and re-run the reverse process. While prior alignment methods focus on what token labels to assign to maximize a target reward, we instead treat which tokens to revisit as the central alignment problem. Selecting edit-positions is challenging because edit effects are interdependent: the impact of modifying one token depends on which others are edited simultaneously. We define an edit-set as a set of token positions to re-mask and re-sample. Motivated by prior work on sparse interactions in biological systems, we find empirically that edit-set value functions for protein inverse folding admit sparse Fourier representations. This structure enables Spectral Feedback to efficiently learn and optimize the value functions for edit-position selection. Spectral Feedback is model-agnostic and can be applied to pretrained, test-time aligned, and fine-tuned diffusion models. For all of these models, the algorithm improves alignment performance without modifying the underlying generative process. Applied to inverse folding with a protein stability reward oracle, it achieves a 32.3% increase in stable proteins for a pretrained model, 24.8% for Best-of-10, and 5.8% for a state-of-the-art RL fine-tuned diffusion model.
Abstract (translated)
URL
https://arxiv.org/abs/2609.30456
PDF
-
TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations
2026-09-24 17:48:20Ayush Jain, Sreeharsha Paruchuri, Ishita Gupta, Fan Zhang, Tanner Schmidt, Jakob Engel, Katerina Fragkiadaki, Adam W. HarleyAbstract
Existing point tracking models face a fundamental tradeoff: they can either track a sparse set of query points over long horizons, or track all points across only short clips. We introduce TrackEverything, a 3D point tracker that breaks this trade-off by representing videos as persistent 3D scene tracks in world coordinates. Grounded in the insight that videos are 2D projections of an underlying 3D world, TrackEverything decouples model complexity from video duration, allowing it to scale with unique physical scene geometry instead. Our approach introduces three key innovations. First, we employ a voxelization-based de-duplication mechanism at sliding-window boundaries to merge co-located tracks, preventing repeated observations of the same surface from redundantly accumulating. Second, we decompose tracking into an endpoint refiner that predicts each point's destination and static-versus-dynamic classification, followed by a lightweight trajectory refiner that decodes dense trajectories exclusively for dynamic points. Third, we propose 3D WAFT, replacing memory-prohibitive 4D correlation volumes with efficient feature sampling in the scene cloud. To the best of our knowledge, TrackEverything is the first 3D tracker capable of tracking all visible points across videos exceeding 1000 frames within 40 GB of GPU memory. On TAPVid-3D, TrackEverything outperforms all open-source all-frame dense 3D trackers by more than 20% APD on short clips, while remaining competitive with state-of-the-art sparse trackers on long sequences, despite tracking far more points.
Abstract (translated)
URL
https://arxiv.org/abs/2609.30222
PDF
-
SAGE: Mitigating Long-Horizon Reasoning Biases via Topological Guidance
2026-09-24 17:33:29Xinyue Zeng, Jiawei Zhang, Yujun Yan, Dawei ZhouAbstract
Long-horizon reasoning remains a central challenge for large language models (LLMs) under sparse-reward regimes. We argue that this brittleness arises from two biases induced by complex reasoning spaces: an exploration bias, where models are drawn toward locally plausible but structurally unstable branches, and a compounding bias, where small local deviations accumulate across depth and suppress rare rewards. We introduce Symbolic Closure Analysis (SCA) as a theoretical lens characterizing how branching structures and sparse rewards induce these biases in long-horizon reasoning with local admissibility, and as a design principle for structural priors in less formal reasoning tasks. Motivated by this analysis, we propose SAGE (Structural Admissibility-Guided Exploration), a unified framework that injects structural guidance to alleviate exploration bias and compounding bias in long-horizon reasoning. SAGE combines two complementary structural guidance: algebraic sparsification, which projects locally admissible candidates onto operator-indexed algebraic subspaces to suppress spurious branching and mitigate exploration bias, and hyperbolic structural guidance, which embeds reasoning states into a negatively curved space to provide dense depth-wise signals and mitigate compounding bias. Across 12 benchmarks and 7 model families, SAGE outperforms competitive baselines. In particular, SAGE achieves up to an 8-fold improvement on the Andrews-Curtis problem, an open real-world long-horizon task. Code is available at: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.30192
PDF
-
Ego-Exo4D Human Meshes Dataset: 4D Human Motion Reconstruction for Ego-Exo Captures
2026-09-24 17:30:36Abhiram Maddukuri, Georgios PavlakosAbstract
Ego-Exo4D is a large-scale dataset providing synchronized egocentric and multi-view exocentric video, a rich resource for skill learning and assessment, procedural activity understanding, and embodied AI. However, the dataset ships with only sparse 3D human pose annotations, and reconstructing dense human motion from its multi-view captures is nontrivial. To this end, we present Ego-Exo4D-HM, a large-scale dataset of 4D human motion reconstructions for Ego-Exo4D's captures, and release the accompanying reconstruction pipeline. The code, dataset, and documentation can be found at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.30187
PDF
-
From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation
2026-09-24 15:34:11Mengdan Zhu, Yufan Zhao, Yao Zhao, Sophie Di, Tao Di, Yulan Yan, Sridhar Iyer, Liang ZhaoAbstract
Semantic IDs (SIDs) encode each catalog item as a short token sequence, enabling generative recommenders to predict the next item autoregressively. Reasoning-enhanced variants, an increasingly common extension, first generate a textual trace and then decode a next-item SID by beam search. Such recommenders are commonly trained with group-relative policy optimization under an exact-match SID reward, which is sparse in large catalogs. Two failure modes follow. When all rollouts in a group miss the target, the group yields zero advantage and no learning signal. Rollouts sharing the same SID reward receive identical advantages, however much their traces differ. In both cases the reward reflects only the decoded SID, never the reasoning that produced it. This creates a credit-assignment gap. We address this gap with retrieval-grounded query attribution. Each trace is structured into a history summary, a set of interest hypotheses, and a final SID. A frozen retriever executes every hypothesis as a catalog query, so that each hypothesis becomes independently verifiable rather than judged only through the final SID. A rollout is rewarded when any of its queries retrieves the target within the \mbox{top-$K$}, and per-query hit indicators localize that reward to individual hypotheses. Credit is thus assigned at the span level: only hypotheses that individually hit receive positive retrieval advantage, while the retrieval channel never updates the final SID span. Rollouts that share a SID reward can therefore receive different updates. Across experiments on three Amazon Reviews datasets, this yields consistent improvements in SID recommendation. On Video Games, an oracle analysis further reveals the potential of interest-conditioned SID decoding: selecting the target-relevant query among generated interests improves both recall and ranking.
Abstract (translated)
URL
https://arxiv.org/abs/2609.29983
PDF
-
Mind What Matters for Reasoning: Aligning Cross-Modal Attention via Selective Probability Mass Concentration
2026-09-24 15:05:22Jiaqi Deng, Zonghan Wu, Zhan Heng, Xiaoshui Huang, Huan Huo, Guandong XuAbstract
Multimodal large language models (MLLMs) achieve strong performance on visual reasoning tasks, yet remain prone to hallucinations and over-reliance on language priors, often generating answers without adequately using task-relevant visual evidence. Existing approaches primarily improve reasoning through reasoning-oriented supervision or inference-time strategies. In this work, we study a complementary question: can multimodal reasoning be improved by strengthening implicit visual grounding without directly supervising the reasoning process? Motivated by the functional specialization of attention heads, we investigate whether reasoning can be improved by guiding only the heads most responsive to visual evidence grounding. We propose Selective Probability Mass Concentration (sPMC), a training framework that identifies grounding-responsive heads and selectively regularizes their text-to-image attention. sPMC treats normalized attention over visual tokens as a spatial probability distribution and encourages the probability mass to be assigned to semantically relevant regions using segmentation-derived spatial priors. Adaptive Head Selection restricts this guidance to visually responsive heads while leaving the remaining heads unconstrained to preserve their complementary functions. Across 6 multimodal benchmark suites, sPMC achieves an average zero-shot improvement of 3% and gains of up to 11.3% across multiple MLLMs while regularizing only 3%-15% of their attention heads. These results demonstrate that targeted guidance of sparse and implicit visual evidence pathways can directly improve multimodal reasoning.
Abstract (translated)
URL
https://arxiv.org/abs/2609.29940
PDF
