Search arXiv⌕ Search

arXiv subjects

Bin Hu

Publications and source records attributed to Bin Hu.

At least 19 recordsLinked to original sources
arXiv2026

UnityBench-Mini3D: Evaluating LLM Agents for 3D Game Development in Unity

Game development stresses coding agents in ways general software engineering does not: correctness depends on how engine APIs, serialized scene state, and runtime behavior interact, not on source code alone. Existing benchmarks mainly target platforms such as Godot and heavily rely on LLM judges or coarse task-level outcomes. We present UnityBench-Mini3D, a benchmark of 77 project-level Unity tasks organized around seven expert-defined skill categories rather than tutorials or genres, with three difficulty tiers. Each task pairs a complete Unity project with natural-language requirements and is evaluated using 1,178 held-out, human-authored milestone tests covering scene structure, component configuration, and runtime behavior. This in-engine evaluation provides deterministic partial credit, localizes failures, and distinguishes partial implementations from complete solutions without relying on subjective judges. Across 21 models and 9 coding harnesses, the average pass rate reaches 80.9% on easy tasks but falls to 35.2% on hard tasks, with just 57.7% overall. UnityBench-Mini3D provides a reproducible, diagnostic testbed distinguishing systems across difficulty levels.

cs.SE↗
arXiv2026

ASPEN: Adaptive Semantic Reasoning via Command Persistence Prediction for Efficient Navigation VLAs

Navigation vision-language-action (VLA) policies repeatedly invoke costly semantic reasoning even when the next command remains unchanged. Command reuse can reduce computation but risks missing instruction-dependent turns or stops. We introduce ASPEN, an event-triggered scheduler based on command persistence prediction: estimating whether the frozen base policy would reissue its previous command. A lightweight multimodal predictor combines temporal observations, language instructions, execution history, and cached VLA confidence. At eligible command boundaries, sufficient predicted persistence permits command reuse. Otherwise, ASPEN triggers fresh semantic reasoning from the VLA. Evaluated with frozen NaVILA on cross-dataset RxR-CE val-unseen, ASPEN reduces VLA calls by 9.5% and estimated inference time by 8.1%, including scheduling overhead. It maintains comparable navigation performance and preserves identical baseline action sequences on 92.1% of episodes.

cs.RO↗
arXiv2026

AV-FABLE: Audio-Visual Manipulation Forensics in Long-Form Videos

Localized AI manipulations in long-form videos often affect brief events within otherwise authentic content, making video-level classification insufficient for forensic analysis. Existing audio-visual deepfake benchmarks largely focus on short talking-head clips, while long-form temporal forensics remains predominantly visual. We introduce AV-FABLE, a dual-timeline benchmark for sparse manipulations in open-scene long-form audio-visual videos. AV-FABLE contains 16,650 pristine--manipulated twin pairs (998 hours) created with 27 editing configurations, with separate audio and visual timelines that enable modality-specific temporal localization and expose errors hidden by joint-timeline evaluation. On AV-FABLE, zero-shot prompting of audio-visual large language models (AV-LLMs) yields near-chance detection and poor temporal localization. In contrast, supervised readouts of their frozen perception features recover strong forensic signals, suggesting manipulation evidence is encoded in these representations but not effectively exposed through language outputs. Motivated by this observation, we propose NFSR(Native Forensics via Separated Readouts), which reads frozen AV-LLM features with separate audio and visual temporal decoders. NFSR achieves the best audio-visual localization performance in-domain and across all three out-of-distribution settings, including transfer to held-out videos from an external long-video forensic benchmark. AV-FABLE, code, and model weights will be publicly released.

cs.CV↗
arXiv2026

Drift Says When, Budget Says How Much: Rate-Controlled KV Re-selection for Long-Form Multimodal Generation

KV cache compression for multimodal LLMs usually selects the visual tokens to keep once, at prefill, and reuses that selection for the whole generation. We show that this one-shot selection is severely limited on long-form generation: on detailed image description, even an oracle that selects with the attention of the uncompressed answer reaches only 0.44-0.52 ROUGE-L agreement with the uncompressed output, because the evidence the model reads moves as the description unfolds. Re-selecting whenever attention drifts out of the kept set is the natural remedy, but we find that a fixed drift threshold fails systematically: drift magnitudes are not comparable across samples, so low-drift samples are almost never refreshed although they benefit the most. We propose Rate-Controlled Refresh (RCR), which lets drift decide when and where to re-select while a per-sample controller regulates how often, combined with a low-rank drift detector, candidate-restricted re-selection, and a decision that stays on the GPU. With one setting for all models and tasks, RCR outperforms the per-step retrieval method InfiniGen by 1.6-3.8 ROUGE-L points on five of six model-task settings with no significant difference on the sixth, while staying within 1.7 points of exact per-step re-selection, with comparable estimated data movement per step and a 4x smaller GPU-resident detector. On DetailCaps, drift-chosen timing beats random timing at the same refresh rate and periodic timing even at a higher rate, by 2.0-5.0 points. On pre-registered video description, where drift is homogeneous across samples, RCR is on par with InfiniGen on Qwen2-VL and exceeds it on Qwen3-VL but no longer beats a fixed threshold at the nearest tested rates, consistent with the diagnosis; object-level evaluation shows that RCR achieves higher object F1 against the uncompressed output than InfiniGen on two of three models.

cs.LG↗
arXiv2026

SAFTER: Safety-Aware Filtering Through Explicit Recoverability for Robust Reinforcement Learning under Distribution Shift

Learning safe policies that generalize to unseen environments is difficult under bounded actuation, where an unsafe decision may become unrecoverable. Conventional control barrier function (CBF) filters enforce only instantaneous safety and ignore future recoverability. We propose SAFTER (Safety-Aware Filtering Through Explicit Recoverability), which embeds a differentiable, closed-form backup-CBF filter into Soft Actor-Critic training. SAFTER unifies current safety, finite-horizon backup safety, terminal recoverability, and input admissibility into a single differentiable certificate. Its CBF condition yields an exact minimum-intervention projection without iterative online optimization. In quadruped navigation, SAFTER reaches 90% training success with 30.6% fewer transitions than Vanilla SAC and 34.1-35.1% fewer than current-state CBF-filtered SAC baselines. Across unseen simulated environments, SAFTER improves success by up to 42.4 and 61.4 percentage points over the state-of-the-art methods CBF-RL and SEA-Nav, respectively, and achieves 73.3% and 76.7% success on two real-world navigation maps. Ablations identify accurate differentiation through the safety filter as a key contributor to robustness, in which replacing the projection Jacobian with the identity approximation reduces success rate in a High-Shift environment from 61.4% to 6.2%. These results show that recoverability-aware filtering serves as both a runtime safety mechanism and a learning mechanism for sample-efficient, robust policies.

eess.SY↗
arXiv2026

PACE-MoE: Adaptive Domain Sampling and Domain-Specific LoRA for Medical Visual Grounding

Medical visual grounding requires models to align clinical descriptions with spatial evidence across heterogeneous imaging domains. However, existing multi-domain post-training strategies typically rely on fixed domain sampling and shared parameter updates, overlooking the need to adapt both which domains receive training and where their updates are applied. To address this issue, we propose PACE-MoE, a policy-guided framework that jointly performs adaptive domain sampling and domain-specific adaptation for medical visual grounding. PACE dynamically assesses the training value of each imaging domain from current-policy feedback and progressively adjusts domain coverage through domain admission, replay, and final all-domain balancing. Meanwhile, domain-specific LoRA adapters route updates to the corresponding imaging domains while preserving the pretrained backbone and visual encoder. This coordinated design aligns training-data allocation with parameter adaptation without modifying the original reward formulation or Group Relative Policy Optimization (GRPO) objective. Experiments across seven medical imaging domains show that PACE-MoE consistently improves grounding performance over fixed-domain training, achieving a 3.64% relative improvement in mean box IoU over the Vanilla baseline. Under matched domain-specific adapter architectures, adaptive sampling alone further improves mean box IoU by 2.07%, demonstrating the benefit of policy-guided domain allocation. These results establish coordinated data-parameter routing as an effective strategy for multi-domain medical visual grounding.

cs.CV↗
arXiv2026

FIT-MAC: A Closed-Loop Multi-Agent Framework for Robust Vision-Language-Action Models

Vision-language-action (VLA) models have achieved strong performance on standard robotic manipulation benchmarks. However, their reliability can degrade substantially when task conditions deviate from the nominal setting. Perturbations in object positions and task instructions may lead to complete task failure. Existing VLA policies generally lack an explicit runtime mechanism for distinguishing these failure modes and selecting an appropriate recovery strategy. To systematically mitigate this issue, we propose FIT-MAC, a failure-aware, closed-loop multi-agent control framework for improving the robustness of VLA policies at inference time. FIT-MAC augments a VLA policy with an inference-time control layer. First, the Planner Agent parses the language instruction into a structured task specification that captures the manipulation goal and spatial constraints, and the Grounder Agent resolves each instruction-level reference to a visual target. Next, the Verifier Agent tracks phase-specific execution progress, distinguishes target-binding errors from local execution failures, and determines whether intervention is warranted. The Correction Agent then routes permitted interventions to target rebinding or bounded grasping or placement recovery, depending on the failure type. When no intervention is required, the pretrained VLA remains the default action generator. Extensive experiments on LIBERO-PRO demonstrate that FIT-MAC improves robustness across multiple VLA backbones, consistently increasing success rates and achieving state-of-the-art performance under diverse perturbations.

cs.RO↗
arXiv2026

RealControlBench: A Benchmark for Interpretability-Informed Fine-Grained Control of LLMs and Agents

Existing steering benchmarks are built around simplified, isolated generation tasks and evaluate a single behavioral attribute under controlled settings that do not reflect the diversity and complexity of practical deployment. As a result, performance on these benchmarks provides only a limited picture of how steering methods perform in the real world. We therefore introduce RealControlBench, a comprehensive benchmark for evaluating fine-grained steering of LLMs and LLM agents across diverse behavioral targets and interaction settings. RealControlBench extends steering evaluation beyond isolated generations to complex, execution-grounded, multi-turn agent interactions, covering five complementary dimensions: instruction-following control, agentic tool-use decision control, reasoning-performance steering, reasoning-length control, and trajectory-conditioned safety awareness. Rather than measuring steering success alone, the benchmark jointly evaluates target controllability and capability preservation, including instruction following, reasoning performance, and downstream agent utility and safety. It further examines whether steering effects remain specific and robust across prompt variations, domains, and model architectures. Our evaluation shows that existing steering methods remain unreliable across realistic control objectives: they often fail to consistently induce the intended behavior and can substantially degrade the model's underlying task performance. These results suggest that strong steering effects in controlled settings do not necessarily translate into reliable control in more realistic evaluations. By providing a unified, model-agnostic steering evaluation framework, RealControlBench aims to support the development of steering methods that are not only effective but also precise, robust, and useful in realistic AI deployments.

cs.CL↗
arXiv2026

DRA-SAM: Disagreement-Driven Referential Active SAM for Low-Altitude Drone Referring Segmentation

Referring segmentation in low-altitude drone imagery requires distinguishing small, visually similar instances under complex linguistic and spatial constraints. Global inference may yield multiple plausible masks or leave the intended referent insufficiently represented, motivating selective local re-inference. We propose DRA-SAM, a disagreement-driven framework that allocates an additional inference pass according to competing referential hypotheses and an unresolved state. A lightweight EvidenceScorer estimates beliefs over existing candidates and the unresolved state. A training-free Referential Disagreement LOOK policy balances belief-weighted candidate competition and unresolved-region coverage to select a crop for local re-inference. The same benchmark-adapted SAM3-I backbone processes the selected region, and the resulting evidence updates existing hypotheses and introduces new candidates. Among the compared methods, DRA-SAM achieves state-of-the-art performance on six of seven test metrics on both RIS-LAD and UAVid-RIS, including mIoU and P@0.5-P@0.9, and on all seven RIS-LAD validation metrics. On RIS-LAD, DRA-SAM improves the matched global SAM3-I baseline from 49.55 to 52.34 mIoU and from 46.80 to 58.05 P@0.5. On UAVid-RIS, the independently adapted DRA-SAM achieves 73.26 mIoU and 86.86 P@0.5. Controlled crop-policy comparisons demonstrate the effectiveness of disagreement-guided crop selection, while target-size analysis reveals particularly pronounced gains for aerial targets occupying less than 1% of an image. These results establish referential disagreement as an effective criterion for allocating local inference. Code is available at https://github.com/phy709/DRA-SAM.

cs.CV↗
arXiv2026

Learning When to Adapt: Reactivity-Conditioned Residual Adaptation under Human-Robot Distribution Shift

Pedestrians differ in whether they respond to a robot, yet many RL policies assume all are non-reactive to prioritize safety. Policies trained without conditioning on local pedestrians' reactivity may therefore be unnecessarily conservative or rely on cooperation that does not occur. We propose RCRA, an online adaptation framework that modulates a learned residual around a frozen conservative navigation policy according to estimated pedestrian reactivity. A Transformer-based estimator infers latent pedestrian reactivity from motion histories and robot context, providing a more accurate signal for navigation adaptation and improving scenario-averaged simulation accuracy by $12.62$ percentage points over the Bayesian baseline, which assumes known pedestrian goals. These estimates are spatially aggregated into an adaptation scale that gates LoRA modules, enabling cooperative behavior near reactive pedestrians and conservative behavior near non-reactive pedestrians. Among baselines achieving SR above $90\%$ in every scenario, RCRA reduces scenario-averaged path length by $16.7\%$ relative to the conservative backbone and by $7.3\%$ relative to randomized-training baselines. It improves mean SR by $14.88$ percentage points over the reactive-trained policy.

cs.RO↗
arXiv2026

Who May Release an Agent's Promise? Standing over Commitments in Multi-Party Chat

When an assistant serving a team makes a promise to Alice, who may cancel it? We call this standing: only the creditor (the person owed) or someone the creditor or operator policy authorizes may release a commitment, regardless of rank. In a controlled chat environment we vary who cancels and separate two errors: unauthorized release (obeying someone without standing) and override (executing a released promise). Across six models, smaller open ones override the creditor (Qwen3.5-9B: 69%); larger ones drop promises on a peer's direct word (Qwen3.8-27B: 89%; gpt-6.1-sol: 83%); nearly all act on a bystander's "go ahead after all". Oracles dissociate the errors: stating the rule or the creditor curbs unauthorized release but not override, which the latter even raises in three of four models; among them, only knowing whether the promise is still owed cuts override. Enabling thinking moves smaller models the way scale does. A self-kept commitment ledger stops override in four of five models; with a one-paragraph release policy it matches or beats the best oracle store, raising gpt-6-luna's strict accuracy from 49% (policy alone) to 83%. Code, prompts, outputs and labels are released.

cs.AI↗
arXiv2026

Lipschitz Bounds for Neural Networks Using a Two-Copy Linear Fractional Transformation

This paper proposes a new method to bound the global Lipschitz constant of a neural network (NN). Our formulation uses a two-copy linear fractional transformation (LFT) representation of the NN with stacked signals. We use biased quadratic constraints (QCs) to bound the input/output behavior of the nonlinear activation functions. We then formulate a linear matrix inequality condition, using these biased QCs, to bound the Lipschitz constant of the NN. The best (smallest) Lipschitz bound is computed via a convex, semidefinite program. There are two key distinctions of the proposed two-copy formulation in comparison to existing methods that rely on a single incremental representation with incremental signals. First, the NN bias terms are eliminated in the standard incremental representation, but appear in our formulation. Second, there is a larger class of QCs that can be expressed for the nonlinearities in our formulation than in the incremental representation. We show that our proposed formulation yields less conservative Lipschitz bounds due to these two differences. We illustrate these points via several simple, motivating examples. We also demonstrate the proposed method using a benchmark problem with a NN controller for an inverted pendulum.

eess.SY↗
arXiv2026

DualAffect: Dual-Path Paralinguistic Integration for Generative Audiovisual Emotion Understanding

Understanding human emotion is essential for multimodal systems that interpret both the content of an interaction and the speaker's expressive state. However, connecting paralinguistic cues to generated answers remains challenging, particularly when vocal expression diverges from visual appearance. We propose DualAffect, a dual-path architecture that integrates visually conditioned acoustic features and temporally organized acoustic tokens into a shared language decoder. The two pathways reuse the same acoustic encoding. The fusion pathway associates sound with visual content through visual queries, while the direct pathway compresses the valid acoustic sequence into ordered tokens through an independent projector. This design gives the decoder access to both joint audiovisual context and acoustic evidence organized along its own timeline. Mixed-task adaptation with task-dependent output preservation supports emotion recognition and general audiovisual question answering within the same model. DualAffect achieves an average metric of 67.89 on CA-MER and accuracies of 62.88% on the EmotionTalk test set and 49.87% on Daily-Omni. Pathway interventions show that the trained decoder uses direct acoustic evidence alongside existing fusion. With model parameters and fusion inputs fixed, zeroing the direct tokens reduces EmotionTalk validation accuracy by 9.75 percentage points and the CA-MER average metric by 2.66 points. The largest CA-MER subset effect occurs when the overall emotion agrees with audio but differs from vision. These findings identify acoustic representation organization as a practical design axis for connecting paralinguistic perception to unified audiovisual generation. Code is available at https://github.com/YiboWang2002/DualAffect.

cs.MM↗
arXiv2026

Empty Commitments: When Agents Promise What They Cannot Deliver

A chatbot that says "I will remind you tomorrow" will not run again until the user writes. We call such a promise an empty commitment: a promise of action after the current turn that nothing in the agent's tools or runtime can carry out. Unlike a broken promise, its emptiness is decided by the agent's configuration at the moment of speaking, so it can be detected from a single turn, before deployment or at run time. We define empty commitments on top of commitment semantics, with three failure types, an anchoring condition for promises that a tool could make real, and an outcome taxonomy that separates these failures from honest deferrals and from over-refusal. We build a checker (a setup-blind detector, deterministic feasibility rules, and a response judge) validated against 400 human labels. On a controlled benchmark of 293 follow-up requests across five setups that add one persistence affordance at a time, four open-weight models of 8-14B parameters fail on 45.9% of responses when no tool exists and nothing is stated. A frontier model fails on 4.4%, but it gets there by deferring and asking, not by using the tools it has: promises made without the enabling call remain in every model. Telling the model its runtime, the cheapest fix, cuts open-weight failures nearly in half where nothing is doable and changes nothing where a scheduler exists; a directive capability card removes most failures at the largest cost in over-refusal; running the checker in the loop and rewriting flagged replies removes more at a smaller cost. Code, prompts, model outputs, and human labels are released.

cs.AI↗
arXiv2026

Evidence for orbital eccentricity supports a hierarchical origin for GW231123

GW231123 is the most massive binary black hole merger observed to date and one of the most intriguing hierarchical merger candidates in the pair-instability mass gap. Its exceptionally large inferred component spins and remarkably short in-band duration make the source highly sensitive to waveform physics near merger. Here we show that GW231123 is better described as an eccentric binary black-hole merger. We analyse data over 11--448 Hz using the eccentric, spin-precessing effective-one-body waveform model TEOBResumS--Dalí, accounting for detector calibration uncertainty. We measure $e_{5.5\,\mathrm{Hz}}=0.34^{+0.12}_{-0.23}$ at 90\% credibility and obtain a Bayes factor of 10.2 in favour of the eccentric model over its quasi-circular counterpart. GW231123-like injection--recovery studies further show that analyses starting at 20 Hz, as in previous studies, can obscure this eccentricity. Including eccentricity substantially changes the spin inference, weakening the preference for a near-extremal secondary spin and lowering its posterior median to $a_2\simeq0.7$, close to the characteristic spin expected for a black-hole merger remnant. These results provide a coherent dynamical interpretation of GW231123, in which residual eccentricity, a mass-gap black hole, and a remnant-like spin arise naturally from a hierarchical merger channel in a dense environment.

astro-ph.HE↗
arXiv2026

FTA-Mem: Fact-Time-Affect Anchored Memory for Low-Density Long-Term Dialogue

Long-term emotional-support agents require memory mechanisms for personalized understanding across sessions. However, emotional-support dialogue is often low-density: turns are incomplete, evidence is scattered, and user states evolve over time. Existing memory methods usually rely on fixed units, such as turn-level notes or session summaries, which may lose details or introduce redundant noise. We propose FTA-Mem, a structured memory framework for low-density long-term dialogue. FTA-Mem uses Boundary-preserving Window Segmentation (BWS) to form coherent situation fragments, and constructs Fact-Time-Affect Memory Units (FTA Units) that jointly encode factual content, temporal grounding, and affective context. Retrieved units are then synthesized into structured context for answer generation. Experiments on ES-MemEval and LoCoMo show that FTA-Mem improves overall long-term memory question answering across benchmarks with different information-density characteristics. On ES-MemEval, FTA-Mem achieves 0.3871 F1 and 0.6668 BERTScore. Further analysis shows that situation-level FTA construction better balances evidence preservation and construction cost than coarse session-level or overly fine-grained turn-pair construction, providing an effective granularity trade-off for long-term dialogue memory.

cs.CL↗
arXiv2026

Adaptive Risk-Certified Event-Triggered Replanning for Dynamic Navigation

Safe navigation in dynamic environments requires robots to plan under obstacle predictions whose errors are uncertain, non-stationary, and can induce rare but safety-critical failures. Existing control-barrier-function safety filters can reject immediately unsafe controls, but they provide little guidance on when the current finite-horizon planning mode itself is becoming unsafe as prediction uncertainty evolves. We propose Conformal Event-Triggered Risk-Certified Replanning (\emph{CERT-Replan}), a framework that uses calibrated barrier risk as an early-warning signal for replanning. CERT-Replan calibrates horizon-indexed obstacle-prediction residuals online and uses the resulting uncertainty radii to evaluate dynamic-obstacle safety margins. A one-step safety filter protects the next applied control, while a horizon-level risk monitor evaluates the upper-tail CVaR of predicted barrier-violation losses along the current MPC rollout. When this risk exceeds an allocated budget, CERT-Replan rejects the current planning mode and selects a lower-risk alternative, such as a different speed profile, corridor, or homotopy class, rather than repeatedly correcting the same nominal plan. In a non-stationary benchmark, CERT-Replan achieves an \(83.3\%\) collision reduction relative to the safety-filter-only baseline \textcolor{black}{and a \(77.8\%\) reduction relative to simple replanning triggers}, while reducing average safety-filter intervention by \(32.4\%\). \textcolor{black}{With Trajectron++, CERT-Replan achieves \(96\%\) collision-free operation. Hardware experiments and onboard runtime profiling demonstrate computational feasibility.}

cs.RO↗
arXiv2026

FC-SWE: Failure-Conditioned RL for Long-Horizon Software Engineering Agents

Repository-level software engineering (SWE) is a challenging long-horizon setting: agents must reason over extended interactions, use tools, and adapt to stateful environments. Recent work trains SWE agents with reinforcement learning methods such as Group Relative Policy Optimization (GRPO), which independently sample multiple trajectories per issue, test the resulting patches, and compare terminal rewards within a fixed group. However, this training setup does not reuse verifier feedback from failed patches as context for subsequent attempts, even though this feedback contains valuable diagnostic information about what went wrong. Training on recovery trajectories is challenging because the preceding outcome determines whether the next trajectory is generated, while the failed execution determines its conditioning context. We introduce FC-SWE, a failure-conditioned RL framework that incorporates recovery attempts into policy training. After a patch fails verification, FC-SWE restores the repository to its original task state and uses the failed patch and verifier feedback as context for a recovery trajectory. FC-SWE adapts GRPO to these chains of complete, multi-turn tool-use trajectories through two mechanisms. Trajectory-local rewards preserve each attempt's verifier outcome, preventing recovery success from rewarding an earlier failed patch. Active-set advantage estimation forms a comparison group from all initial and recovery trajectories actually executed for the same issue, so failed attempts remain in the group while unexecuted attempts are excluded. On all 500 SWE-bench Verified tasks under a verifier-assisted protocol, FC-SWE with Qwen3.5-4B and SWE-agent achieves 41.7% Resolved@1 and 52.8% Resolved@2, compared with 38.9% and 48.5% for GRPO. Although trained with at most two attempts per chain, FC-SWE reaches 70.7% Resolved@11 under an eleven-attempt test-time budget.

cs.LG↗