arXiv subjects

Bing Li

Publications and source records attributed to Bing Li.

At least 19 recordsLinked to original sources

arXiv2026

Nonlinear Sufficient Dimension Reduction for Causal Inference with Continuous Treatments

Causal inference with continuous treatments requires estimating nuisance functions to adjust for the potentially high-dimensional confounder, but nonparametric estimation of these functions is severely hampered by the curse of dimensionality. To address this challenge, we propose a nonlinear sufficient dimension reduction framework to construct low-dimensional sufficient confounders to preserve the unbiasedness of the treatment effect. We estimate the causal effect by combining the generalized Sliced Inverse Regression with the inverse probability weighting, outcome regression, and augmented inverse probability weighting, respectively. We further establish the convergence rates of the first two estimators and derive the asymptotic distributions of the third. Our theoretical results show that, compared to estimators based on the full covariates, dimension reduction yields faster convergence rates and weaker rate conditions without inflating the asymptotic variance. Simulation studies demonstrate that the proposed estimators improve finite-sample performance over their counterparts without dimension reduction, particularly when the original covariate dimension is high. These results demonstrate that sufficient dimension reduction provides a flexible approach to mitigating the curse of dimensionality in causal inference with continuous treatments.

stat.ME↗

arXiv2026

KV Packet: Recomputation-Free Context-Independent KV Caching for LLMs

Large Language Models (LLMs) rely heavily on Key-Value (KV) caching to minimize inference latency. However, standard KV caches are context-dependent: reusing a cached document in a new context requires recomputing KV states to account for shifts in attention distribution. Existing solutions such as CacheBlend, EPIC, and SAM-KV mitigate this issue by selectively recomputing a subset of tokens; however, they still incur non-negligible computational overhead (FLOPs) and increased Time-to-First-Token (TTFT) latency. In this paper, we propose KV Packet, a recomputation-free cache reuse framework that treats cached documents as immutable "packets" wrapped in light-weight trainable soft-token adapters, which are trained via self-supervised distillation to bridge context discontinuities. Experiments on Llama-3.1 and Qwen-3 demonstrate that the proposed KV Packet method achieves near-zero FLOPs and lower TTFT than recomputation-based baselines, while retaining F1 scores comparable to those of the full recomputation baseline. Code and experimental results are available at https://github.com/ChuangtaoChen-TUM/KVPacket.

cs.LG↗

arXiv2026

Chiral magnetic excitations in the RMn6Sn6 kagome metals (R=Tb, Ho, Er, and Y)

Recent inelastic neutron scattering (INS) investigations of the kagome metal TbMn6Sn6 reveal novel chiral magnetic excitations associated with the kagome layer. These excitations appear to have an itinerant character with a large energy scale (> 100 meV) and heavy damping; however, the origin of these excitations, which are not predicted by local moment linear spin-wave theory, is not understood. Here, we use INS to show that chiral excitations are also found in RMn6Sn6 compounds with R=Tb, Ho, Er, and Y having uniaxial, tilted, planar ferrimagnet, and planar double-spiral magnetic structures, respectively. Thus, the chiral excitations exist regardless of the magnetic structure. Additionally, R=Er and Y demonstrate that the chiral mode intensity is weaker in planar magnetic structures, suggesting that excitations are transverse to the local Mn moment direction. The chiral excitations appear only at L=odd positions in the Brilloun zone, and correspond to chiral spin exciations of the same handedness within each kagome layer, and pi out-of-phase between layers.

cond-mat.str-el↗

arXiv2026

HyWe: Unifying Search and Operation for Real-World Hybrid Web Tasks

Real-world hybrid web tasks need both information gathering and page interaction, yet search-based and operation-based agents are developed and evaluated separately. To bridge this gap, we introduce HyWe, a framework that unifies both capabilities within one model and trajectory. Its two modes use dedicated prompts and tool sets to sustain sub-goal execution; explicit switching selects the needed capability, and shared history lets each mode build on the other's findings. To provide demonstrations of these behaviors, we develop HyWeCollector, a fully automatic pipeline that uses task prototypes, four switching patterns, and multi-agent synthesis to produce verified hybrid trajectories. We train HyWe-35B on these demonstrations to unify efficient information gathering and browser interaction in a single 35B web agent. To evaluate hybrid task completion, we introduce HyWeBench: 100 human-written tasks on websites distinct from training, with human-written sub-goal checklists and an outcome-based evaluation pipeline. Hybrid training enables HyWe-35B to achieve 47.0% success on this benchmark, improving on the untrained base model by 19.0 percentage points and outperforming existing search-based and operation-based agents. Even before training, the same base model solves 28.0% of tasks with HyWe, versus 11.0% with all tools under one prompt. Beyond hybrid tasks, HyWe-35B achieves 59.0% and 88.3% success on Online-Mind2Web and WebVoyager, outperforming existing open-source baselines and cutting action cost per successful task by 48.3% and 51.9% compared with Fara-7B.

cs.AI↗

arXiv2026

AeroIntent: A Synchronized Multimodal Benchmark for UAV Mission Intent Recognition

Reliable recognition of mission-level UAV intent from multimodal observations is essential for low-altitude airspace surveillance and security. Existing multimodal UAV datasets primarily support detection, tracking, and localization, leaving limited support for evaluating mission-intent recognition across held-out trajectory programs and incomplete radar histories. We introduce AeroIntent, a simulation-based benchmark comprising 990 synchronized multimodal sequences generated by 22 trajectory programs across five mission intents. The benchmark provides radar, RGB, infrared, and depth observations, together with program-disjoint, radar-gap, and observation-prefix evaluation protocols. Building on this benchmark, we propose AeroMIR, a structured Multimodal Intent Reasoner that integrates radar trajectories, derived behavioral features, selected visual evidence, and task knowledge for candidate-based UAV intent recognition. Under a strict program-disjoint protocol evaluating unseen tactical realizations, AeroMIR achieves a mean accuracy of 71.56% and a mean Macro-F1 of 69.84% across three seeds, outperforming the strongest trajectory-only baseline by 7.79 Macro-F1 points and a conventional radar-RGB late-fusion baseline by 19.32 points. AeroMIR also demonstrates strong performance under controlled radar gaps and early intent recognition settings. AeroIntent and AeroMIR provide a reproducible foundation for multimodal UAV intent recognition. The benchmark is available at: https://anonymous.4open.science/r/AeroIntent

cs.RO↗

arXiv2026

Self-Distillation for Supervised Fine-Tuning

Supervised fine-tuning treats the expert demonstration or gold response as a one-hot target, suppressing valid alternative continuations that the model would otherwise assign probability to. We propose Gold-Conditioned Distillation (GCD), a method that turns expert demonstrations into model-native supervision. GCD uses one model as both teacher and student: the teacher conditions on a prompt containing the gold response in context, the student conditions on the original prompt as at inference, and the student is trained to minimize the forward KL divergence to the teacher's next-token distribution at each gold response token. This lets the model use its prior knowledge to teach itself the new task: correctness is anchored by the gold response, while the soft target preserves plausible alternatives. Across various training tasks on Qwen and Llama models, GCD improves new-task accuracy over SFT and generally reduces degradation on general benchmarks.

cs.CL↗

arXiv2026

Shrinking targets versus recurrence: A survey on the Hausdorff dimension theory

This survey reviews results on the Hausdorff dimension of shrinking target sets and recurrence sets in dynamical systems. In their basic formulations, shrinking target sets concern orbits approaching a prescribed centre, whereas recurrence sets concern orbits returning close to their initial points. A central theme is the comparison of the Hausdorff dimensions of these two classes of limsup sets. The dynamical systems considered include expanding conformal systems, irrational rotations, affine iterated function systems, and matrix transformations. Related results on Hausdorff measure are also included when they provide natural refinements of the dimension theory. Finally, we review results on badly approximating orbits and non-recurrent sets, with particular attention to winning properties and intersections with self-conformal and self-affine sets.

math.DS↗

arXiv2026

Can Failed Trajectories Reveal Where Agents Go Wrong? Agentic Failure Attribution Without Step-Level Labels

Failure attribution for LLM agents aims to identify decisive steps responsible for task failure, which is critical for debugging and improving agentic systems. Existing methods rely on costly LLM prompting, step-level error annotations, or verified successful trajectories that lack failure signals for assistance. In this paper, we study failure attribution from step-unlabeled failure trajectories, where each trajectory is known to have failed but its decisive step is unknown. We propose SOAP (Spectral scOring with Attention-guided Propagation), an effective framework requiring neither step-level training labels nor a separate success-only reference set. SOAP first constructs a spectral reference space from LLM representations and assigns each step a base error score according to its projection behavior. Importantly, later steps may exhibit stronger failure signals because they inherit an earlier mistake, causing independent scorers to favor downstream consequences over the decisive error. To address this downstream contamination, SOAP derives soft step dependencies from attention patterns and propagates downstream error evidence toward its likely upstream sources. Extensive experiments show that SOAP consistently improves attribution performance over competitive baselines.

cs.LG↗

arXiv2026

Constructing Pareto Frontiers of Optimal Reasoning Termination Methods for Overthinking in Reasoning Models

Reasoning models can benefit from test-time scaling, but continued reasoning may become ineffective or harmful, a phenomenon termed overthinking. Evaluations against full thinking compare accuracy and token usage but lack references for overthinking severity and strategies' theoretical mitigation potential. To characterize overthinking at the model-task level, we first introduce Step-wise Truncation Probing (STP), repeatedly generating answers from multiple retained prefixes along each fixed sampled reasoning trajectory. Using each trajectory's earliest measured position attaining maximum empirical accuracy as a reference, we define Overthinking Ratio and Accuracy Degradation as the fraction of thinking tokens generated afterward and the accuracy decrease from this reference to full thinking, respectively. Secondly, we then formulate dynamic routing, fixed budgets, and dynamic early exit within a unified framework to quantify the theoretical mitigation potential of different termination strategies. Based on this unified framework, we use the complete STP measurements to construct a token-accuracy Pareto frontier for each strategy under a common normalized thinking-token budget. We further develop OTBench, a reusable benchmark built from the STP process, which produced 70.5B reasoning tokens and the corresponding probing measurements across six models and four benchmark groups.

cs.AI↗

arXiv2026

The Thinking Leakage in Reinforcement Learning for LLMs

Hybrid reasoning models such as the Qwen3 series switch between a THINK mode and a NO-THINK mode through a few chat-template tokens. A popular recipe trains such models with reinforcement learning in NO-THINK mode and reports improved reasoning performance. We provide evidence that much of this gain can be explained by thinking leakage: the gain comes largely from partially reactivating the reasoning abilities of the original model's THINK mode instead of learning enhanced reasoning behavior. We train Qwen3-1.7B, 4B, and 8B with GRPO in NO-THINK mode. As accuracy improves, responses become longer and more reflective, are increasingly judged by the original model to resemble its own thinking, and the model's internal activations shift toward its THINK mode. We then investigate at the level of weights, activations, and attention heads: scaling the RL update further elicits THINK-mode behavior and approaches THINK-mode accuracy but does not exceed it; in Qwen3-8B, replacing a low-dimensional THINK-related activation subspace of the RL-trained model with its original-model value removes most of the RL gain, and the reverse exchange transfers a considerable fraction of the gain to the original model; and the RL-induced shift passes through the attention heads that read the empty think block. We further identify a possible mechanism in the GRPO training signal: responses favored by the advantage-weighted update are more THINK-like, and a gradient step computed from this signal moves the model toward its THINK state. Thus, gains in NO-THINK mode can largely reflect reactivation of reasoning abilities already present in the original model.

cs.CL↗

arXiv2026

Event-Centric Memory with Query-Aware Graph Augmentation for Long-Term Conversational Agents

For persistent and personalized conversational agents, memory systems can enable them to remember, update, and reason over long histories by storing past interactions and retrieving relevant information. Existing memory systems typically follow two paradigms: flat-structured memory and graph-based memory. The former is lightweight but leaves event relations and state updates implicit, while the latter explicitly models memory structure but incurs additional construction cost and introduces irrelevant relations over long histories. To address these limitations, we propose QGMem, a novel memory construction and activation framework motivated by human memory, in which experience is organized into events and query-relevant events are modeled by graph as working memory. QGMem converts long dialogue histories into event-indexed atomic memory units that preserve individual experiences and consolidates related units into dynamic memory traces that retain state trajectories and current states. When a query arrives, hybrid memory retrieval gathers complementary candidate memories, and query-aware reranking activates the most relevant units as a compact working memory. To expose relational dependencies in the working memory and support conflict-aware reasoning, QGMem organizes the working memory as a local graph, which is then encoded as a graph token and provided to the LLM together with the textual working memory to improve evidence utilization during answer generation. Experiments across six benchmarks validate the framework and show consistent gains in retrieval, multi-hop evidence composition, conflict resolution, and ultra-long dialogue reasoning with compact contexts and moderate inference cost.

cs.CL↗

arXiv2026

Extreme superluminal expansion traces an ultra-narrow GRB jet across relativistic regimes

The long-term evolution of gamma-ray burst (GRB) jets from ultra-relativistic launch to a slowly expanding radio source remains a fundamental problem in high-energy astrophysics, rarely resolved directly by imaging. We use very long baseline interferometry (VLBI) to trace the radio angular scale of GRB 221009A, the brightest GRB ever detected, from 5 to 299 days after the burst, and constrain the jet's geometry and dynamics. The Very Long Baseline Array data at 8.4 and 15.2 GHz are calibrated and imaged with AIPS and Difmap, and the source size is characterized by a circular-Gaussian model fitted to the complex visibilities. The jet dynamics, multi-wavelength emission, and source size are modeled within a two-shell collision scenario. A hydrodynamic simulation confirms that such an ultra-narrow jet undergoes lateral spreading. The observations reveal apparent superluminal expansion, with the expansion speed declining from about 90c at 5 days to about 6c at 300 days, tracing the jet's deceleration from the ultra-relativistic to the mildly relativistic regime and placing GRB 221009A far beyond the luminosity--velocity envelope of active galactic nuclei and microquasar jets. Combined with the multi-wavelength afterglow, they establish an ultra-narrow relativistic jet with a half-opening angle of about 0.01--0.03 radians, propagating through a wind-like medium at small radii before entering an interstellar-medium-like profile at larger radii. Joint modeling favors a two-phase size evolution with tentative evidence for lateral spreading. These results provide a direct geometric view of a GRB outflow and establish GRB 221009A as a benchmark for relativistic jet physics.

astro-ph.HE↗

arXiv2026

Self-Generated Feedback Destabilizes Test-Time Training: A Causal Decomposition of Long-Horizon Adaptation

Test-time training (TTT) lets a model store information in its weights during inference. When the model learns from its own output, however, each update also changes the model that generates the next training example. Across 128K-token streams, retaining generated-text updates worsens prediction on independent human-written text with three TTT-E2E model configurations (labeled 125M, 760M, and 3B). The same failure occurs when Adam updates Qwen3-4B's existing weights. The same update mechanisms can improve on real text, so writing itself is not the failure. Three matched comparisons trace the causal pathway. Fixed Generation removes over 98% of the damage at 125M and 760M by using a frozen model to generate training chunks. Recorded Replay separates the loss caused by reading degraded text from the additional loss stored by updating on it. A paired one-update comparison then shows the local conflict: an update predicts its source better but new real text worse. This cost grows after Closed Loop adaptation, with a few trajectories accounting for most large failures. Finally, Settlement evaluates the candidate state on independent real text before commitment. It leaves mean endpoint gaps of 0.07 and -0.02 nats at 125M and 760M while retaining real-text adaptation. These results motivate checking prediction on independent evidence before retaining an update.

cs.CL↗

arXiv2026

MultiLoc: Look Around As You Localize for Fast And Robust Visual Re-localization

Relative camera pose is a basic geometric cue for visual re-localization and scene understanding. When using images alone, visual evidence may not sufficiently support reliable estimation of 3D-consistent motion. The challenge is especially acute in unfamiliar and ever-changing environments for real-time applications, where a system must infer spatial structure under tight time constraints from what it sees rather than rely on scene-specific reconstruction or training. This raises a fundamental question: how can live 3D camera motion be estimated accurately for re-localization in unseen environments while drawing on enough scene context to resolve pose ambiguity? We introduce MultiLoc, a multi-view-guided relative pose regressor trained at scale to achieve spatial and geometric consistent representations for robust visual re-localization in unseen scenarios with high inference speed. Specifically, MultiLoc creates a minimal 3D-sub-scene representation of the environment and efficiently transforms it into 3D spatially and geometrically consistent features in a single forward pass, enabling precise pose estimates with high inference speed gains. Across diverse indoor, outdoor, and in-the-wild visual re-localization benchmarks---Indoor6, Cambridge Landmarks and WaySpots---MultiLoc consistently outperforms state-of-the-art relative pose regression methods. We also highlight that MultiLoc, a pose regressor, also performs competitively with inference-heavy and structure-based approaches while retaining sub-second inference and generalizing to unseen scenes. Given a small posed support set, it also surpasses relative pose regression, feature-matching, and non-regression methods on several relative camera pose benchmarks. Code will be released.

cs.CV↗

arXiv2026

Lang3DSeg: Annotation-Free Open-Vocabulary 3D Segmentation with Point Transformers

Accurate 3D semantic perception is critical for safe autonomous navigation. However, supervised LiDAR segmentation remains tied to closed taxonomies and to the cost of point-wise manual annotation. Open-vocabulary methods avoid that cost by projecting the output of 2D vision-language models onto LiDAR and distilling it into a 3D network. These methods rely almost exclusively on voxel-based sparse convolutions, and point transformers have so far been limited to indoor environments, where 3D data is dense and bounded. We present Lang3DSeg, which establishes a point transformer as the backbone for annotation-free open-vocabulary segmentation of outdoor 3D LiDAR, and is trained from scratch without geometric pre-training. This training paradigm necessitates addressing the inherent noise in 2D-to-3D label projections; specifically, naive projection often suffers from depth ambiguity, where points behind an object are erroneously assigned its semantic label. We therefore composite masks using an explicit class-priority rule and truncate each projected instance at the first gap in its depth distribution, correcting the projection error directly rather than averaging it over registered sequences. Lang3DSeg achieves 52.8% mIoU on nuScenes validation and 41.4% on SemanticKITTI, the highest among published annotation-free methods on both benchmarks. Every 3D semantic segmentation is on a single LiDAR sweep, and inference operates in real-time without running vision-language models.

cs.CV↗

arXiv2026

Frame Differential On-Policy Self-Distillation for Video Reasoning

Reinforcement learning (RL) has substantially improved the reasoning ability of multimodal language models through verifiable rewards and increasingly fine-grainedvisual or temporal credit assignment. In video reasoning, however, current RL methods typically train with a fixed sparse frame budget: increasing the number of frames makes autoregressive rollouts expensive, while too few frames may miss temporally localized events and fine-grained visual details. We present \textbf{Frame Differential On-Policy Self-Distillation (FD-OPSD)}, which transfers the useful evidence of dense frame observations to a sparse frame policy during RL training. FD-OPSD compares the policy's token level preferences for the same sampled response under sparse and dense views, and distills the resulting frame differential signal without an external teacher or dense autoregressive rollout. The method preserves sparse-frame rollouts and leaves inference unchanged. Across Qwen2.5-VL-7B and Qwen3-VL-4B on six video reasoning benchmarks, FD-OPSD yields higher overall average performance than the strongest corresponding GRPO, T-GRPO, or Video-KTR baselines across the 16, 32, and 64 frame evaluation settings. These results show that dense visual evidence can be transferred selectively during training through token level self-distillation while retaining sparse frame rollouts and unchanged inference.

cs.CV↗

arXiv2026

Lights, Camera, Attack: Exploiting Temporal HDR Fusion with Pulsed Light

Modern cameras widely use temporal High Dynamic Range (HDR) to improve visibility by capturing a sequence of exposures with different integration times and fusing them into a single image. This process implicitly assumes that scene illumination remains sufficiently stable during capture. We introduce FLASH (Fusion-Level Attack by Saturating HDR), an external pulsed-light attack that deliberately attacks this assumption by creating cross-exposure inconsistency before downstream perception. FLASH exploits an algorithmic assumption rather than relying on sensor damage or hardware failure, and requires neither physical camera access, access to raw exposure brackets, knowledge of the fusion algorithm, nor exact phase lock to the camera. Across eight physical camera platforms spanning embedded, surveillance, photography, smartphone, and automotive use cases, and matched optical controls, FLASH causes pipeline-dependent darkening, overexposure, and visibility loss. This includes extreme-darkening rates of 50.0% on an iPhone 16 Pro and 33.7% on a Wyze Battery Cam Pro. On the Wyze camera, FLASH triggers the system-level low-visibility response in 10/10 trials, compared with 0/10 continuous-light and randomized-frequency flashing controls. In a controlled stationary OpenPilot case study, 23.0% of frames exhibit severe darkening in the traffic-cone target region, with target-background CNR decreasing by up to 90.8%. Under FLASH, the OpenPilot interface also fails to display the system-level path state observed in the corresponding control trials. In a controlled night-only HDR reconstruction stress test, a proof-of-concept exposure-rejection defense reduces median output-brightness deviation by 79.16%. These results show that temporal HDR fusion itself requires security-aware validation of exposure evidence.

cs.CR↗

arXiv2026

Structured Interaction, Visual Localization, and Robust Execution for Complex Web Tasks: A Technical Report on the WebRetriever Challenge

This report presents the web agent system developed for the WebRetriever Challenge. The system follows a structuredinteraction- first strategy, using semantic webpage information for routine browser operations and invoking visual perception only when structured representations are insufficient. Three key designs are introduced: grid-assisted visual localization for difficult-to-access controls, hierarchical context management for reducing redundant page and interaction history, and fault-aware execution mechanisms for stable multi-browser task processing. The system achieved a pass rate of up to 79% in local evaluation on Protocol 1. In the official Protocol 3 competition, it achieved a 59% pass rate with eight concurrent browser workers and ranked first overall, winning the WebRetriever Challenge.

cs.IR↗