Searcharxiv⌕ Search

arXiv subjects

Ling Li

Publications and source records attributed to Ling Li.

At least 19 recordsLinked to original sources
arXiv2026

HiMA-Ecom: A Long-Horizon Memory Benchmark for Hierarchical Multi-Agent E-commerce Assistants

Long-horizon agentic systems become long-context systems by accumulation rather than through a single long document, since every step reassembles a prompt from role instructions, tool and sub-agent descriptions, recalled memories, dialogue history, and previous tool outputs, separately for every agent. This paper studies how such accumulated context can be trained, evaluated, and managed in hierarchical multi-agent assistants. We present HiMA-Ecom, a hierarchical multi-agent benchmark for e-commerce assistants with 22.8K instances, of which 17.7K carry memory entries in context, pairing agent-specific supervised fine-tuning data with system-level supervision for end-to-end multi-agent reinforcement learning. Building upon it, HiMA-R1 keeps the context budget of a trajectory under control through Variance-Reduction Group Relative Policy Optimization (VR-GRPO), which samples along an initial trajectory so that sampling cost grows additively with trajectory length and updates the agents with the largest reward variance, together with an efficiency reward on trajectory length and a memory evolution mechanism that repurposes GRPO rewards as cost-free signals for what stays in context. Agents built upon 3B and 7B open-source models match DeepSeek-R1 and surpass DeepSeek-V3 by 6.0 points on average, while reward-driven memory reduces the update steps needed for convergence by roughly 55%.

cs.AI↗
arXiv2026

Correlation-Aware Exploration for Long-Trace Reasoning

Reasoning models trained with Group Relative Policy Optimization (GRPO) now produce chains of thought thousands of tokens long, and recent studies reveal that such training is prone to policy collapse. Entropy regularization is the standard remedy, but it works by pushing probability mass away from the few tokens that carry it and toward the vast low-probability tail of the vocabulary, and it applies this force once for every token of the reasoning trace. This puts the bonus in direct conflict with the entropy decay that convergence requires, and makes its coefficient difficult to set in the long-trace regime. To resolve this tension, we propose Info-GRPO, an information-theoretic framework that cultivates correlation between the policy and a latent prior. Rather than raising entropy everywhere, Info-GRPO builds on the mutual information between a latent variable and the trajectory it induces: augmenting prompts with latent seeds during training lets the model explore diverse policies correlated with the prior, while the seed-conditioned entropy is guided toward convergence. Diversity is therefore carried across whole reasoning traces rather than by per-token randomness, and exploration no longer fights entropy reduction. Experiments show that Info-GRPO outperforms vanilla GRPO and entropy-regularized GRPO across diverse reasoning benchmarks. On AIME24, it improves $Avg@8$ over GRPO by 3.75%, 1.66%, 4.16%, and 4.17% on four backbones, and it reaches the same majority-vote accuracy with $8\times$ fewer samples. Code and models will be publicly released.

cs.LG↗
arXiv2026

Frontier-OPD: Discovering the Edge of Agent Capability and Teaching Agents to Go Beyond

On-policy distillation aligns teacher feedback with a student's current behavior, yet acquiring useful supervision for multi-turn agents requires identifying both where guidance is needed and whether the student can sustain the interaction afterward. We introduce Frontier-OPD, which addresses this problem through the student's learning frontier: decision situations that remain challenging but admit useful autonomous continuation after brief teacher guidance. To operationalize this frontier, Frontier-OPD organizes on-policy experience into a shared-state trajectory graph, where trajectory-balanced aggregation of teacher-student disagreement identifies distinct intervention candidates while preserving the histories needed for replay. Short teacher bridges are then assessed by comparing agreement over the bridge and the ensuing student continuation with an equally long segment of the original trajectory, providing a behavioral proxy for guidance that the current student can build on. Accepted teacher responses augment ordinary distillation data, and fresh rollouts renew the frontier estimate as the policy changes. Experiments on ALFWorld, WebShop, and ScienceWorld report higher mean autonomous performance than the compared distillation baselines across the evaluated student scales and teacher configurations, with ablations indicating complementary effects of candidate selection and guidance filtering. In one ALFWorld timing setting, the additional collection cost is offset by a shorter time to a target success rate.

cs.CL↗
arXiv2026

HarnessRL: Co-Evolving Policies with a Just-in-Time Training Harness for Open-Ended RL

Rubric-based reinforcement learning (RL) trains policies for open-ended generation with natural-language criteria as rewards, but these rewards can become ineffective as training progresses. We identify two failure modes. Under non-discriminative rewards, scores collapse to uniformly low or high values due to insufficient exploration or a mismatch between task difficulty and policy capability. Under spurious rewards, higher scores do not correspond to better responses. Both reflect a training configuration whose task sampling, exploration, and evaluation stay fixed as the policy evolves. We introduce HarnessRL, a training-time harness that co-evolves with the policy through three interfaces for task sampling, rollout guidance, and rubric criteria. HarnessRL maintains a training-time memory from which a harness evolver retrieves within-task history and cross-task evidence. Using this evidence, the evolver attributes each reward failure to the interface whose update can address it. It then runs a proposer and just-in-time critic loop in which the critic accepts a candidate update only after measuring its effect on the current frozen policy, returning rejected candidates for revision. We also curate Rubric-Mix-10k, a five-domain training and evaluation suite for open-ended RL across medical, science, writing, role-play, and instruction-following tasks. On Rubric-Mix-10k, HarnessRL outperforms the strongest baselines by 2.8 to 3.1 points across both 4B and 9B model scales and nearly halves the training steps to its best checkpoint. Code and data are available at https://github.com/Tongyi-ConvAI/Qwen-Character/tree/main/HarnessRL.

cs.CL↗
arXiv2026

Reclaiming Low-Quality Instruction Data with Rule-Guided LLM Reconstruction

Data selection for supervised fine-tuning (SFT) often discards low-scoring instruction examples that may retain useful supervision. We propose a rule-guided LLM reconstruction framework to reclaim this information. The framework identifies low-scoring samples through LLM-based rating and score correction, selects semantically representative sources, and reconstructs instruction-response pairs through strategy-guided generation, criteria-based validation, and failure-guided repair. From 123,786 low-scoring samples, it constructs 54,888 training examples. Across three 7B/8B base LLMs, the reconstructed data achieves higher average downstream performance than the low-scoring source set and all evaluated data-selection baselines. Despite using substantially fewer examples than the full source dataset (approximately 300K), it achieves higher reported average scores than full-data training across all three base models. Stepwise ablations identify rule-guided reconstruction as the largest source of observed improvements over direct fusion, while a manual audit reveals remaining limitations in source integration and answer faithfulness. These findings support reconstruction as a complementary approach to selection-based SFT data curation.

cs.CL↗
arXiv2026

What Makes a Social Agent's Interactions Effective? Learning from Verifiable Outcomes in Paired Worlds

Effective social interaction requires more than fluent responses: agents must reach plans that fit the participants' needs and circumstances. Existing evaluations of social agents mainly rely on LLM judges that rate the conversation, without checking the resulting plan. We introduce HICO-Bench, a benchmark of 180 everyday coordination scenarios, such as resource allocation, scheduling, and authorization, in which two agents with private information converse and then independently record the plan they agreed on. A program checks Feasibility (the two records match and satisfy both participants' private conditions), and an LLM judge checks Integrity (the conversation supports the recorded plan); passing both is a Verified Success. Each scenario has two versions that share the same public context but differ in one private condition, so the correct plan differs between them. Across thirteen LLMs, 36% of conversations rated highly for goal completion fail Verified Success. To learn from outcomes that arrive only at the end of a conversation, we propose Paired-Credit Policy Optimization (PaCPO). After each turn, PaCPO asks the agent for a tentative plan, checks it in both versions of the scenario, and rewards the turns after which the tentative plans fit their respective versions. PaCPO raises Qwen3-8B's Verified Success from 20.4% to 28.3%.

cs.CL↗
arXiv2026

S$^3$Geo: Structure-Semantic Synergistic Learning for Cross-View Geo-Localization

Cross-view geo-localization (CVGL) aims to estimate geographic locations by matching images captured from different viewpoints, such as drone and satellite views. Existing methods mainly rely on visual representations, but often fail to jointly model fine-grained structural correspondences and semantic priors, making them prone to confusion between visually similar but semantically different regions, and thus limiting robustness under large viewpoint variations. To address these challenges, we propose \textbf{S$^3$Geo}, a structure-semantic synergistic learning framework for cross-view matching. Specifically, we first introduce a Decoupled Query Pooling (DQP) module to extract a compact set of region-aware features from dense tokens, enabling explicit modeling of local structural patterns. We then design a query-level contrastive learning scheme with an optimal transport (OT)-based formulation to establish soft correspondences under cross-view spatial misalignment. Furthermore, we incorporate a Semantic Knowledge Distillation (SKD) strategy from a frozen CLIP teacher to transfer semantic priors and relational structures, thereby improving discrimination on hard negatives. By operating synergistically, the semantic priors provide robust contextual filtering, which guides the structural module to establish precise spatial alignments. Experiments on the University-1652 and SUES-200 datasets demonstrate that \textbf{S$^3$Geo} consistently outperforms state-of-the-art approaches without increasing inference complexity, validating the effectiveness of jointly modeling structural and semantic information for CVGL.

cs.CV↗
arXiv2026

Spatial Latent Reasoning for Embodied Reference Understanding

Pointing-gesture visual grounding requires connecting hand geometry with the visual identity and extent of a referred object. A central challenge for continuous latent reasoning is how to organize these complementary cues into useful intermediate supervision. We propose Spatial Latent Reasoning (SLR), a framework that structures this supervision around an ordered sequence of geometric and visual states. A spatial ray state is supervised by fingertip position and pointing direction, followed by four states aligned with target-region features. To construct the visual targets, we introduce parity pooling, which applies polyphase grouping to average region tokens on four interleaved spatial supports. All states are generated recurrently during training and inference; auxiliary annotations are required only during training. On EgoPoint-Ground, the framework improves mIoU over same-backbone supervised fine-tuning by 2.8, 17.5, and 21.1 percentage points on Qwen3.5-4B, Qwen2.5-VL-7B, and Qwen3-VL-8B, respectively, with improvements on both hard subsets. On YouRefIt, it achieves 77.6% precision at IoU 0.5, a numerical margin of 5.2 percentage points over the reported state of the art under differing evaluation protocols. Ablations support joint geometric and visual supervision on the standard and similar-object sets, and favor parity over three alternative pooling operators on the standard set. These results support task-structured supervision for continuous pointing grounding. We will release the code and supporting materials.

cs.CV↗
arXiv2026

SkillCycle: Co-Evolving Agent Policies and Skill Banks

Internalizing external skills changes a language agent's capabilities and, with them, the value of its remaining guidance: rules can become redundant, misleading, or insufficient for newly encountered decisions. This creates a coupled problem of learning from skills and adapting the skills that supervise further learning. We introduce SkillCycle, a framework for co-evolving agent policies and skill banks through a feedback loop between skill internalization and rule revision. Our central contribution is to give distillation feedback a second role: token-level contextual differences help locate rules for inspection, while interaction outcomes guide edits to their content and applicability. SkillCycle alternates between two phases: policy learning with a fixed skill bank and router, and rule revision with a frozen policy. Candidate edits undergo rule-level and whole-bank environment comparisons before they guide the next learning cycle. On WebShop, SkillCycle with a 3B model achieves a success rate of 74.74% and a score of 88.37 without inference-time skill inputs, representing relative improvements of 0.73% and 3.96% over the state-of-the-art (SOTA) model, respectively. In Cycle 3 ablations on ALFWorld and WebShop, SkillCycle's no-skill success rates improve by 10.18% and 18.11% relative to a static skill bank, and by 2.41% and 2.50% relative to a single bank update, respectively. These results show that continually revising skill guidance as the agent's capabilities change helps transform external skills into policy capabilities that require no skill inputs at inference. We will release code, configurations, skill banks, and evaluation protocols.

cs.CV↗
arXiv2026

Timestep Weighting: A Hidden Key to Effective ELBO-Based Flow-Matching RL

ELBO-based reinforcement learning offers a sampler-agnostic approach to fine-tuning flow matching models with reward feedback. Timestep weighting in ELBO-based RL has large impact on performance, and it also provides a unified view (as we show in this work) to understand prediction losses heuristically chosen in prior work, yet it remains under-researched and is often chosen to inherit pretrain configs. We investigate impacts and dynamics of timestep weighting in ELBO-based RL. We show that effective weighting depends on both the reward landscape and stage of learning. (1) Through experiments on controlled CIFAR image generation, complemented by robotics, we investigate how weighting impacts reward-driven updates across noise levels. (2) Through gradient analysis, we reveal distinct patterns of cross-noise coordination across tasks and their evolution during training. These findings motivate the hypothesis that useful weighting depends on the gap between the policy's current behavior and the behavior favored by the reward. (3) Guided by this analysis, we study simple static weighting, budgeted profile selection, and dynamic schedules that improve performance beyond conventional target choices. Our results establish timestep weighting as an important design choice for flow-matching RL and motivate further research into methods that choose and adapt it throughout learning.

cs.LG↗
arXiv2026

VISTA: Internalizing Collective Visual Experience via On-Policy Distillation for Active Multimodal Agents

Active multimodal agents use visual tools to acquire task-relevant evidence while reasoning. Although reinforcement learning samples multiple interaction trajectories per input, outcome-based objectives primarily use the group to estimate scalar advantages, leaving complementary visual discoveries underused. We introduce VISTA, which internalizes collective visual experience through on-policy distillation by turning observations from same-input rollouts into shared supervision. Collective visual experience distillation (CVED) organizes these observations with their interaction context and aligns them with individual decisions, while heterogeneity-aware policy improvement (HAPI) reinforces successful trajectories and provides experience-guided distillation for unsuccessful attempts. An experience-conditioned teacher evaluates the student's sampled response prefixes, allowing discoveries from one trajectory to guide learning in another without replacing the student's original history or generating new target trajectories. The trained agent retains its visual tools and acts using its own interaction history. VISTA achieves the strongest average performance among the evaluated active multimodal agents of comparable size and consistently outperforms same-backbone training baselines across fine-grained perception and general reasoning tasks, demonstrating the value of collective experience for active multimodal learning.

cs.CV↗
arXiv2026

CODEBLOCK: Learning to Supervise Code at the Right Granularity

Supervised fine-tuning of code LLMs typically applies uniform cross-entropy loss to all response tokens, implicitly assuming that every token provides equally useful learning signals. Recent token-level selection methods challenge this assumption in natural-language SFT by supervising only high-value tokens. However, such pointwise selection can fragment the syntactic structures and program dependencies of code, leaving supervision scattered across incomplete code units. In experiments on 30K code instruction-response pairs, under the same 10% token budget, supervising complete coding blocks improves average performance by 11.7 points over prior approaches that supervise isolated tokens. Motivated by this observation, we propose CodeBlock, a structure-aware sparse supervision framework that uses complete, parser-aligned coding items as the basic units of supervision. CodeBlock constructs coding items from high-quality code instruction data, estimates their supervision utility using GCE, which is more robust to low-probability tokens, and further adjusts their supervision priority using data-flow reach and bridge signals. During training, the full response is retained as context, while loss is applied only to the selected supervision units. Experiments show that partial supervision over only about 6.9% of response tokens consistently outperforms full-token SFT across all five model settings, suggesting that effective code SFT depends not only on identifying high-value supervision, but also on allocating it at the appropriate structural granularity.

cs.LG↗
arXiv2026

AUV-Bench: Aesthetic Understanding and Generation Evaluation for User Interfaces

Multimodal foundation models are increasingly used for evaluating and generating user interfaces (UIs), often producing seemingly reasonable aesthetic judgments and visually plausible pages. However, under professional design scrutiny, their behavior can differ substantially from that of human designers. In professional design practice, designers rely on a systematic set of aesthetic principles that consistently guide judgment, diagnosis, repair, and creation. A coherent aesthetic capability should therefore connect aesthetic judgment with design actions. Existing evaluations, however, typically assess these abilities in isolation, making it difficult to determine whether task-level success reflects a shared aesthetic understanding or merely fragmented task-specific competence. To address this gap, we introduce AUV-Bench, developed in collaboration with professional UI designers around 1,395 executable web interfaces and four tasks: aesthetic scoring, diagnosis, repair, and text-to-UI generation. The tasks share a pool of UIs and aesthetic principles, with diagnosis and repair further aligned on 660 controlled-degradation instances to enable instance-level analysis of judgment and action. Evaluation of 12 models reveals a capability imbalance: models show moderate agreement with professional designers in holistic aesthetic scoring, yet exact diagnosis-chain success peaks at only 24.7%. On the aligned diagnosis-repair cases, correct judgments and successful repairs do not consistently coincide, exposing a Judgment-Action Gap between identifying aesthetic problems and successfully acting on them. In open-ended generation, even leading models achieve only moderate aesthetic quality under human-calibrated evaluation. Overall, current models exhibit partial aesthetic competence, but still lack the fine-grained understanding and judgment-action coherence required for reliable UI design.

cs.AI↗
arXiv2026

KernelZero: Co-Evolving Proposer and Coder for Continuously Improved GPU Kernel Generation

High-performance GPU kernels are essential to modern machine learning systems, yet automatically generating kernels that are both correct and efficient remains challenging. Existing LLM-based approaches face two major limitations: the scarcity of high-quality training data aligned with the model's current capabilities, and the inherent trade-off between kernel correctness and performance. To address these challenges, we propose KernelZero, a co-evolution framework that continuously improves GPU kernel generation through two specialized models: a Proposer that generates Torch modules from API sets and a Coder that translates them into CUDA or Triton kernels. KernelZero uses a frontier-driven module generation mechanism to continuously produce capability-aligned training modules based on the Coder's current weaknesses. It further introduces Correctness-Aware Group Relative Policy Optimization (CA-GRPO), which optimizes performance only after correctness becomes sufficiently reliable. By alternating the optimization of the Proposer and Coder, KernelZero forms an automatic curriculum that enables targeted and training-efficient capability improvement. Empirically, KernelZero-7B surpasses Claude-4.5-Sonnet on CUDA and DeepSeek-V4-Pro on Triton. On KernelBench Level 1 and 2, it achieves CUDA pass@1 scores of 75.8% and 69.6%, respectively, with pass@10 reaching 100% and 97%. On Triton, it achieves pass@1 scores of 77.2% and 72.5%, respectively.

cs.LG↗
arXiv2026

Towards Whole-Study Screening for Congenital Heart Disease in Fetal Ultrasound Using Multiple Instance Learning

Congenital heart disease (CHD) is the most common birth defect, yet a large fraction of cases remain undetected on prenatal ultrasound, in part because current artificial-intelligence methods assume that the key diagnostic frames have already been isolated from a study, by a clinician or by a view classifier. We remove that assumption and address CHD screening directly at the level of the whole ultrasound study. We propose a two-stage framework that first learns transferable frame representations by self-supervised masked-autoencoder pre-training on unlabeled fetal ultrasound, then identifies cardiac frames with a disease-robust module and aggregates them with a transformer-based multiple instance learning (MIL) model that produces a case-level diagnosis from study-level labels alone. The model further returns its highest-scoring frames for clinician review, and a hierarchical head separates critical from non-critical CHD. On the internal test set of our multi-source development cohort (FUSE), the proposed cardiac-gated MIL model reaches an area under the curve (AUC) of 0.985 with a specificity of 0.990, outperforming the reproduced NATMED ensemble (AUC 0.861, specificity 0.600) and the FetalCLIP foundation model (AUC 0.867, specificity 0.710). On an independent external cohort, all models initially perform near chance, but label-free CORAL adaptation raises the proposed model from an AUC of 0.513 to 0.944, whereas whole-study and view-dependent baselines do not recover. These results indicate that whole-study MIL with disease-robust cardiac-frame identification is an accurate and deployable route to prenatal CHD screening.

eess.IV↗
arXiv2026

ZigZag Trie: A Novel Index for Contextual Queries

There is increasing interest in queries about the context of a string $P$ in a longer text $T$, i.e., the set of all string pairs $(L,R)$, with $|L|=|R|=q$, for a given $q$, such that the string $LPR$ occurs in $T$. Such contextual queries are important in several domains but are challenging to answer efficiently. This is because the length of $T$ in applications is massive and existing indexes do not directly encode the context of a given $P$, which is key for answering retrieval queries efficiently. Our work introduces the ZigZag Trie (ZZT), a new full-text index to specifically address these challenges. This index reorganizes the text so that, for any $P$, all possible strings $L$ and $R$ growing symmetrically around $P$ are grouped into a common subtree of the index, allowing their efficient retrieval. We show how to construct the ZZT of $T$, which has size $\mathcal{O}(n)$ where $n=|T|$, in $\mathcal{O}(n\log n)$ time and $\mathcal{O}(n)$ space. On top of ZZT, we design specialized indexes that, for a query pattern $P$, answer four new types of contextual queries: (I) finding the longest string $LPR$ that occurs at least $τ$ times in $T$, for a fixed $τ$; (II) finding the longest string $LPR$ that occurs in at least $τ$ texts of a text collection, for a fixed $τ$; (III) reporting the total number of distinct contexts of $P$ in $T$; and (IV) retrieving, for a given $q$, the $k$ pairs $(L,R)$ of $P$ with the highest scores according to a given scoring function. Our indexes answer queries of type I, II, and III in optimal time, and of type IV in near-optimal time. Moreover, their size, construction space, and construction time are linear or near-linear in $n$, given ZZT. Using real billion-letter datasets, we show that our indexes answer queries orders of magnitude faster than baselines and perform similarly or better in index size and construction space and time.

cs.DS↗
arXiv2026

Multiflagellarity facilitates bacterial upstream motility

Upstream swimming drives bacterial spreading and surface colonization. Many pathogens encounter fluid flows as they infect the intestines, lungs, and urinary tract, so how bacteria use their flagella to counter these flows matters for disease and treatment. Yet how morphology and flagellar arrangement govern motility against flow remains unknown. Here, we investigate the biophysical determinants of rheotaxis by combining microfluidics, directed evolution, genetics, holography, and hydrodynamics simulations. Using upstream swimming competitions, we find that peritrichous E. coli and S. enterica rapidly outcompete monotrichous P. aeruginosa and V. cholerae, accumulating upstream at densities up to five orders of magnitude higher, even though Vibrio swims three times as fast. Motility selection experiments show that rheotaxis increases with flagellar number and length, confirmed by overexpressing the master regulator flhD/C. Three-dimensional holography and single-cell tracking reveal that multiflagellarity stabilizes surface residence and promotes the weathervane effect that reorients cells upstream, a mechanism further supported by simulations that fully resolve flagellar arrangement and fluid-structure interactions. These results establish multiflagellarity as a key facilitator of upstream navigation, governed by near-wall residence and shear-driven reorientation rather than by swimming speed.

physics.bio-ph↗
arXiv2026

High-Performance Scaled P-Type SnOx Transistor by Atomic Layer Deposition with CFET Integration

The development of high-performance p-type oxide semiconductors is essential for realizing complementary logic for monolithic 3D integration, yet p-type oxide semiconductors still exhibit substantially inferior performance compared with their n-type counterparts. In this work, we demonstrate high-performance p-type SnOx transistors by atomic layer deposition (ALD), as back-end-of-line compatible devices for monolithic 3D integration. The SnOx transistors exhibit high field-effect mobility of 6.9 cm2/Vs, low subthreshold swing (SS) of 185 mV/dec, decent on/off ratio (ION/IOFF) of 1.8*104 and high bias stability. By scaling the channel length down to 80 nm, a high on-current of 38.7 mA/mm at VDS of -1 V is achieved. It is understood that precursor and reaction engineering to suppress Sn4+ component in SnOx film are the key for performance enhancement. Furthermore, a complementary field-effect transistor with ALD SnOx p-FET vertically stacking on ALD In2O3 n-FET is also demonstrated, achieving maximum voltage gain of 21 V/V at VDD of 4 V. These findings suggest ALD SnOx as a promising candidate for scaled high-performance BEOL p-type transistors.

cond-mat.mtrl-sci↗