arXiv subjects

Miao Li

Publications and source records attributed to Miao Li.

At least 19 recordsLinked to original sources

arXiv2026

DecompGym: Benchmarking LLM Agents on Designing Decomposition Algorithms

Benders decomposition and column generation solve large optimization problems by passing cuts, columns, and dual values between a master problem and subproblems, and are hard to implement correctly. Large language models now automate parts of algorithm design, which points toward agents that build decomposition solvers for new problems. As a first step toward such agents, we study how well current coding agents design and implement these two algorithms, and how they fail. We introduce DecompGym, 81 instances from nine problem families (five Benders, four column generation), in which an agent writes a solver for a prescribed method. Each solver must meet three criteria: execution within a 60-second time limit; numerical feasibility together with a method audit, in which an LLM judge reads the solver's code; and an objective within 1% of the stored optimal objective. We evaluate 12 GPT, Claude, and Gemini models running in coding-agent tools. Over up to ten attempts, we compare independent regeneration with reflection, which first diagnoses the failure. On the first attempt, two models solve every instance and seven solve more than 90%. Most Benders failures (48 of 62) are missing or suboptimal solutions. The most common column-generation failure is a method-audit rejection. Output checks alone would accept 93 of the 108 attempts that fail the method audit. The most common rejection, an optimalif, occurs 96 times, 92 of them in one model.Reflection turns 66% of retries into successes, compared with 39% for regeneration, and it repairs method-audit failures, for which the judge names the defecective gaps. Agents for decomposition therefore need checks on how a solver reaches its answer and feedback that locates the error. The benchmark is available at https://github.com/SIKAI-C/DecompGym.

cs.AI↗

arXiv2026

CRIM: Composable and Reusable Item Memory for Efficient Personalized Retrieval

Large language models incur substantial training and inference costs in personalized search and recommendation, as growing interaction histories lengthen request inputs and repeatedly expose the same item text to the model. We propose \textbf{C}omposable and \textbf{R}eusable \textbf{I}tem \textbf{M}emory (CRIM), a framework that separates item-level semantic encoding from request-level history modeling. A small set of learnable summary tokens compresses each item into compact continuous memories, while segment-level readout tokens (REC tokens) introduce predictive supervision within the interaction history. These memories are independent of users and queries and can therefore be reused across requests. At inference time, CRIM composes cached memories in interaction order and incorporates an optional query to produce a personalized retrieval representation, reducing repeated item-text encoding while retaining request-dependent contextual modeling. We further introduce a joint training mechanism that combines cached memories with sampled live re-encoding. Re-encoding only a few history items per request allows the retrieval objective to jointly optimize the item and request encoders at limited computational cost, while periodic cache refreshes maintain consistency. Extensive evaluation and cost analysis on MovieLens-1M, Qilin, and our new LazSeq dataset show that CRIM achieves state-of-the-art retrieval performance with substantially lower inference costs at the same backbone scale, while limiting item re-encoding during joint training. We will publicly release our code and dataset.

cs.IR↗

arXiv2026

AgentPhishing: A Benchmark of LLM Email Agents Under Adaptive Attack

Computer-use agents increasingly handle private information and act on users' behalf across desktop and web applications, including email. Phishing therefore poses privacy and security risks that depend not only on whether agents recognize malicious messages, but also on how they respond. Existing phishing datasets were developed primarily for conventional detection, leaving agent susceptibility to persuasive malicious requests insufficiently characterized. To expose vulnerabilities that pre-existing datasets may miss, we introduce AgentPhishing, a benchmark for evaluating whether persuasive phishing attacks can evade detection and induce LLM email agents to comply with malicious requests. We develop PhishOpt, which optimizes reusable email-generation skills against individual agents under fixed benchmark constraints and a realism constraint. We analyze the generated emails using Cialdini's persuasion principles to identify strategies associated with agent failures and use these findings to construct a universal generation skill. Our evaluation shows that agents can handle pre-existing phishing emails safely yet remain susceptible to attacks adapted to agent behavior, while in an AgentPhishing evaluation of nine agents, four exhibit phishing-detection rates below $10\%$ and malicious-request compliance rates above $45\%$. These results highlight the limits of inferring agent safety from conventional phishing evaluations and the need to assess how agents respond to persuasive malicious requests.

cs.CR↗

arXiv2026

Task-Level Natural Language Priors as Learning Signals for Low-Resource LLM Training

Large language models (LLMs) often struggle when low-resource training data are ambiguous or incomplete. Task-level natural-language priors can provide useful guidance in such settings, but existing approaches usually treat these priors as input context rather than as learning signals during training. We study how task-level natural-language priors can supply guidance missing from these data through Prior-Guided Tuning (PGT). Its implementation, Contrastive Prior Steering (CPS), retains the supervised objective and adds positive- and negative-prior-conditioned losses with opposing optimization directions. Across AmbiMath, Diagnostic Anchoring, and Jigsaw, CPS achieves near-perfect exact-match accuracy on the synthetic task, improves accuracy under misleading diagnostic suggestions by 8.2 percentage points, and improves average subgroup Macro F1 on Jigsaw by 10.7 percentage points over plain fine-tuning, averaged over the two evaluated models. On Jigsaw, with one tenth of the experimental training data, CPS slightly exceeds full-data plain fine-tuning. These results show how reusable task-level language can complement limited labeled examples without requiring instance-level rationales or preference pairs.

cs.AI↗

arXiv2026

Beyond Context Length: Characterising the Difficulty of Long-Context Reasoning

Long-context language models support ever longer context windows, and evaluations typically treat context length as the main indicator of difficulty. However, context length alone cannot explain why models fail on long-context tasks that require reasoning over complex information. Contexts of equal length can pose very different challenges: a task may be solvable from a single local passage, or it may require integrating evidence distributed across the full context. We quantitatively characterise this difficulty through three properties: evidence scope, how much necessary information must be found; evidence dispersion, how widely that information is scattered within the context; and reasoning depth, how many dependent steps are needed to derive an answer from it. We construct LURCH, a multi-domain question-answering benchmark of 6,550 instances with supporting evidence and reasoning graphs, enabling all three properties to be quantified. Evaluations of proprietary and open-weight models show that: (1) at fixed context length, all our three properties are strongly negatively correlated with model performance; (2) as context length increases, performance degrades sharply on samples with dispersed evidence, whereas it remains relatively stable on samples with localised evidence; and (3) difficulty compounds across properties, as questions with moderately high values on all three are harder than those with an extreme value on just one. These findings establish evidence scope, dispersion and reasoning depth as measures of long-context reasoning difficulty that complement context length.

cs.CL↗

arXiv2026

Target-Token Asymmetry Can Masquerade As Depth Specialization

Layer-ablation studies often perturb early or late layers of a language model, compare the damage to prediction on two corpora, and read a larger early-layer effect on one corpus as early-layer specialization for its task. Such a comparison averages over the tokens that each corpus asks the model to predict, and these target tokens differ: math solutions, for example, ask for far more whitespace and numbers than encyclopedia prose. We show that this difference can create the appearance of depth specialization. We split a cross-corpus early--late damage interaction by target-token type and recompute it under a target mixture shared by both corpora. Between GSM8K and Wikitext, a positive interaction replicates across five model instances. In the three fully decomposed models, only whitespace has a positive interaction. Removing whitespace reverses the pooled effect in all three, and matching the full target mixture reverses it in two. The diagnosis holds under position matching, adversarial reassignment of token types, and a continuous readout. The risk can also be anticipated from the tokenizer alone. A screen computed before any model runs places every held-out pair it scored in the region its label predicts, and on the pairs it rates at high or medium risk, removing whitespace reverses every positive pooled estimate. A smaller content-word effect one token ahead follows the share of content targets that the unablated model already predicts. A cross-corpus depth effect therefore supports an interpretation of task specialization only after the target-token mixture is reported, the effect is decomposed by token type, and a mixture-matched summary keeps its sign.

cs.CL↗

arXiv2026

Match the Canvas, Not Just the Answer: Predictability-Aware SFT for Diffusion Language Models

Diffusion language models (DLMs) enable flexible parallel generation by resolving tokens according to the model's evolving predictability rather than a fixed left-to-right order. This flexibility, however, introduces training choices absent from standard autoregressive teacher forcing: which reference tokens are visible when a target is supervised and how much generation space is allocated. Standard supervised fine-tuning (SFT) typically makes these choices through random masking, independently of inference. Our analysis shows how this training-inference mismatch can misdirect supervision. Random masking can conceal prediction errors or penalize uncertainty that decoding would otherwise resolve, while the attention-visible input length can change which response formulations the model favors, even when the underlying solution is unchanged. Motivated by these findings, we propose Predictability-Aware Canvas Selection and Supervision (PACS), an SFT method. PACS uses the model's own predictability to select each example's attention-visible input length and supervises reference tokens immediately before their confidence-guided revelation. Across reasoning and code-generation benchmarks, PACS consistently achieves the highest mean accuracy among the compared SFT methods, with gains of up to 12.84 percentage points over vanilla SFT on individual benchmarks. Code will be released upon publication.

cs.LG↗

arXiv2026

Controllable and Photorealistic Pedestrian Risky Motion Generation for End-to-End Driving Safety Evaluation

Evaluating end-to-end autonomous driving under rare, safety-critical vehicle-pedestrian interactions requires photorealistic, sensor-level scenarios. However, trajectory-based scenario generators cannot synthesize raw visual observations, whereas video-based approaches lack controllability. To bridge this gap, we present ControlPed, a novel framework that combines trajectory-level conflict synthesis with 3D Gaussian Splatting (3DGS) to generate photorealistic, motion-controllable safety-critical scenarios. Built upon HazardPed, a dataset derived from 10,352 traffic videos comprising 422 conflict trajectories, HD maps, and 857 annotated 3D human motions, ControlPed first generates conflict trajectories, lifts them into 3D human motion sequences via text-conditioned motion diffusion, and finally renders multi-view sensor observations using animatable 3DGS avatars. Safety evaluation in 88 rendered photorealistic scenarios reveals that seven leading end-to-end driving models suffer a severe performance drop, with their mean HDScore plunging from 88.8 to 47.4, exposing major failure modes under dangerous pedestrian behaviors. The dataset and testing benchmarks will be released to facilitate safety assessment of vehicle-pedestrian interactions.

cs.RO↗

arXiv2026

Exploiting Vulnerabilities: Universal Adversarial Attacks on Vision-Language-Action Models in Robotics

Recently, Vision-Language-Action (VLA) models have revolutionized robotic manipulation by seamlessly integrating visual perception, language understanding, and action generation in an end-to-end learning framework. However, since these models are designed to interact directly with the physical world and humans, their security is critical, and even small vulnerabilities can lead to catastrophic failures. In this work, we propose the Universal Adversarial Object, a sphere with optimized surface texture that significantly degrades task success rates when placed within the robot's field of view. Specifically, our approach introduces a multi-level attack framework that jointly disrupts trajectory planning, task execution, and action control. We validate our method in both simulated and real-world robotic settings. Experimental results demonstrate that the adversarial object reduces the average task success rates by 31.2%-39.9% for two representative VLA models (Pi0 and RDT), with success rates dropping to near zero in complex scenarios. Index Terms--Vision-Language-Action models, adversarial attack, robotic security, universal adversarial object

cs.RO↗

arXiv2026

Simple Agentic Memory for Generalist Robot Policies

Visual-memory systems commonly retain or compress past observations. Robot control additionally requires interaction-derived state that no individual frame may explicitly represent, such as persistent identity relations, accumulated progress, or ordered procedures. We introduce Simple Agentic Robot Memory (SimpleARM), a training-free memory layer for frozen generalist robot policies. From the task instruction, SimpleARM specifies what to monitor; frozen perceptual tools maintain compact typed state online; structured access retrieves that state only when a proposed subgoal depends on history; and current-view grounding resolves recalled entities before execution. We evaluate SimpleARM on RoboMME, a benchmark of memory-dependent robot manipulation tasks that require history information no longer available in the current observation. Across all 16 tasks and three policy seeds, SimpleARM achieves 67.17% mean success, compared with 44.51% for the strongest non-oracle baseline. Matched ablations show mechanism specificity: removing relation, reference, progress, or route state produces large losses where the affected state is retrieved for control, while largely sparing other tasks. These results support a state-based view of robot memory: effective memory for control is not simply retained visual history, but compact task-relevant state derived from the interaction history.

cs.RO↗

arXiv2026

ICI-VLA: In-Context Imitation with Spatiotemporally Aligned Demonstrations for Vision-Language-Action Models

Vision-Language-Action (VLA) policies are commonly adapted to new manipulation settings through additional gradient updates, which limits rapid deployment when task-specific data or compute is scarce. We present ICI-VLA, a training and retrieval framework that equips a text-action VLM with few-shot test-time adaptation through in-context demonstrations. Unlike mainstream VLA designs based on action-specific multimodal fusion, ICI-VLA retains the native text-generation interface. ICI-VLA updates its parameters only during offline training; at inference, the policy remains fixed and conditions action generation on retrieved micro-demonstrations. The framework decomposes long trajectories into short, semantically labeled examples and trains an RD-Encoder with positives mined by Dynamic Time Warping (DTW), aligning the retrieved context with the phase and geometry of the current subtask. We further introduce Target Action Masking, a context-corruption objective designed to reduce direct action copying and increase reliance on the current observation. ICI-VLA reaches average success rates of 97.7% on LIBERO and 60.4% on RoboTwin 2.0, exceeding the highest reported baseline average on RoboTwin 2.0 by 19.3 percentage points. It also achieves 83.2% across four physical tasks. These results indicate that a fixed VLA policy can benefit from conditioning on spatiotemporally aligned demonstrations at test time.

cs.RO↗

arXiv2026

SciTrek: Evaluating and Improving Long-Context Numerical Reasoning over Scientific Articles

We introduce SciTrek, a synthetic question-answering dataset for assessing and improving long-context numerical reasoning in large language models (LLMs). Existing long-context datasets with inputs beyond 64K tokens either target simple information retrieval or, when they do involve reasoning, rely on artificial contexts, while numerical reasoning remains largely overlooked in both cases. SciTrek addresses these limitations with questions that require numerical operations (e.g., counting, sorting, aggregation, and comparison) over collections of full-text scientific articles. Questions are generated automatically by formulating them as SQL queries over a database of article metadata (titles, authors, and references), and ground-truth answers are obtained by executing these queries. The underlying SQL provides a transparent, verifiable specification of the reasoning each question requires, enabling fine-grained error analysis. Furthermore, the proposed fully automated pipeline scales to arbitrary context lengths and dataset sizes with minimal human supervision, mitigating data contamination and supplying abundant data for post-training. Extensive experiments show that frontier open-weight and proprietary LLMs struggle even with ostensibly simple questions: the best-performing model achieves only 46.5% exact match at 128K tokens, and performance degrades steadily as contexts grow. Fine-grained analysis reveals systematic weaknesses on citation-related questions and on compound logical conditions, particularly those involving negation. Finally, post-training open-weight models on SciTrek improves their numerical reasoning in ways that generalise to out-of-domain long-context tasks.

cs.AI↗

arXiv2026

Pay for The Second-Best Service: A Game-Theoretic Approach Against Dishonest LLM Providers

The widespread adoption of Large Language Models (LLMs) through Application Programming Interfaces (APIs) induces a critical vulnerability: the potential for dishonest manipulation by service providers. This manipulation can manifest in various forms, such as secretly substituting a proclaimed high-performance model with a low-cost alternative, or inflating responses with meaningless tokens to increase billing. This work tackles the issue through the lens of algorithmic game theory and mechanism design. We are the first to propose a formal economic model for a realistic user-provider ecosystem, where a user can iteratively delegate $T$ queries to multiple model providers, and providers can engage in a range of strategic behaviors. As our central contribution, we prove that for a continuous strategy space and any $ε\in(0,\frac12)$, there exists an approximate incentive-compatible mechanism with an additive approximation ratio of $O(T^{1-ε}\log T)$, and a guaranteed quasi-linear second-best user utility. We also prove an impossibility result, stating that no mechanism can guarantee an expected user utility that is asymptotically better than our mechanism. Furthermore, we demonstrate the effectiveness of our mechanism in simulation experiments with real-world API settings.

cs.GT↗

arXiv2026

Fishing Out Free Riders: Shapley-Based Reward Attribution for Parallel Reasoning via Reinforcement Learning

Large Language Models (LLMs) excel at multi-step reasoning, yet current parallel reasoning approaches often fail to distinguish the contributions of individual reasoning paths. Many paths may be redundant, misleading, or even detrimental, but outcome-level rewards assign uniform reward, leading to ambiguous learning signals and unstable training. We propose Parallel Shapley, a reinforcement learning framework that attributes fine-grained, path-level contributions in multi-path reasoning. Treating each path as a player in a cooperative game, we leverage Shapley values to quantify marginal contributions, using a generative reward model to evaluate path utilities and Monte Carlo sampling for efficient approximation. Experiments on mathematical reasoning benchmarks show that Parallel Shapley outperforms existing baselines while providing more stable and interpretable training. Our framework effectively "fishes out the free riders," assigning reward proportionally and improving multi-path reasoning in LLMs.

cs.AI↗

arXiv2026

Dynamic High-frequency Convolution for Infrared Small Target Detection

Infrared small targets are typically tiny and locally salient, which belong to high-frequency components (HFCs) in images. Single-frame infrared small target (SIRST) detection is challenging, since there are many HFCs along with targets, such as bright corners, broken clouds, and other clutters. Current learning-based methods rely on the powerful capabilities of deep networks, but neglect explicit modeling and discriminative representation learning of various HFCs, which is important to distinguish targets from other HFCs. To address the aforementioned issues, we propose a dynamic high-frequency convolution (DHiF) to translate the discriminative modeling process into the generation of a dynamic local filter bank. Especially, DHiF is sensitive to HFCs, owing to the dynamic parameters of its generated filters being symmetrically adjusted within a zero-centered range according to Fourier transformation properties. Combining with standard convolution operations, DHiF can adaptively and dynamically process different HFC regions and capture their distinctive grayscale variation characteristics for discriminative representation learning. DHiF functions as a drop-in replacement for standard convolution and can be used in arbitrary SIRST detection networks without significant decrease in computational efficiency. To validate the effectiveness of our DHiF, we conducted extensive experiments across different SIRST detection networks on real-scene datasets. Compared to other state-of-the-art convolution operations, DHiF exhibits superior detection performance with promising improvement. Codes are available at https://github.com/TinaLRJ/DHiF.

cs.CV↗

arXiv2026

Co-policy: Responsive Human-Robot Co-Creation for Musical Performances

Art has long stood as a pivotal expression of human creativity. Embodied artificial intelligence offers a route for generative models to participate in that creativity through physical action rather than disembodied digital content. In robotic music co-creation, it is challenging to connect semantic musical understanding with real-time and physically executable performance. We present Co-policy, a framework for human-robot musical co-creation that separates semantic intent grounding, constrained musical variation, and visuomotor execution. To ground musical semantics, Co-policy uses pre-inference semantic anchors and a fine-tuned Qwen-vl planner (F-Qwen) to transform speech, live musical seeds, and visual observations into structured co-creation plans. To support low-latency execution, Co-policy introduces a Gaussian-Mixture Visuomotor Policy (GMP), implemented as a conditional mixture-density policy that maps target notes and visual context to multimodal robot actions in a single forward pass. Unlike robotic playback systems that merely reproduce user-specified notes, Co-policy generates complementary musical responses under both musical and physical constraints. Real-robot chime experiments, ablations, and expert evaluation show improved intent alignment, execution accuracy, and response frequency over diffusion-policy and ablated baselines, supporting physically grounded action generation as a key requirement for embodied human-AI co-creation.

cs.RO↗

arXiv2026

SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

Agent Skills are structured packages of procedural knowledge that augment large language model (LLM) agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark whose current inventory contains 87 tasks across 8 domains paired with curated Skills and deterministic verifiers. Our latest aggregate evaluation runs the 87-task benchmark under matched no-Skills and curated-Skills conditions for 18 model-harness configurations. Curated Skills raise the average pass rate from 33.9% to 50.5% (+16.6 percentage points; 25.5% normalized gain), with configuration-level gains ranging from +4.1 to +25.7 pp. Focused Skills with at most three modules outperform larger or exhaustive bundles, and smaller models with Skills can match larger models without them. SkillsBench establishes paired evaluation as the foundation for rigorous measurement of Skill efficacy on agentic, expertise-heavy work.

cs.AI↗

arXiv2026

ReportQA: QA-Based Radiology Report Evaluation

Radiology report evaluation is essential for advancing automated report generation. Natural language generation metrics have limited clinical relevance. Clinical efficacy (CE) metrics evaluate important medical findings, but focus mainly on presence and cover only a limited set of entities. Due to heavy reliance on manual annotations, it is difficult for CE metrics to extend clinical entities or attributes. In clinical practice, radiology reports serve as a medium for information transfer. Clinicians use them to perform downstream diagnostic tasks without directly inspecting images. Based on this insight, we propose ReportQA, a clinical-related and flexible radiology report evaluation framework, supporting detailed quantitative analysis of radiology report generation systems. We first collect datasets covering multiple imaging modalities and anatomical regions. We then construct knowledge trees of clinical entities and attributes with radiologist guidance, and use large language models (LLMs) to extract structured information from raw reports. Next, we generate QA pairs from predefined templates and apply quality control through self-filtering and report-based filtering. During evaluation, the report is treated as context, and an LLM acts as a judge model to answer the QA pairs. Based on the resulting QA accuracy, we introduce QAScore metric. Compared with existing metrics, QAScore shows better alignment with radiologist judgments. Experiments on multiple state-of-the-art vision-language models reveal that current report-based inference paradigms struggle to learn fine-grained clinical representations and exhibit strong negative prior biases. In contrast, question-driven inference provides a more effective alternative. For reproducibility and extensibility, we release the knowledge trees, structured reports, and QA pairs, along with the pipeline code for QA construction and evaluation.

cs.CL↗