arXiv subjects

Ke Zhang

Publications and source records attributed to Ke Zhang.

At least 19 recordsLinked to original sources

arXiv2026

LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes

Sparse Mixture-of-Experts (MoE) models offer scalable capacity for Web applications, but uneven routing creates computation and communication bottlenecks. Existing placement and replication methods redistribute this workload, yet additional expert copies compete with token-related buffers for limited GPU memory. We present LRMoE, a memory-aware MoE training system that jointly recalibrates backward demand and balances the retained workload. During forward execution, it plans retention quotas and expert placement under resource constraints while preserving all routed computations. During backpropagation, a load recalibration metric estimates each token-expert pair's contribution to the expert update and selects pairs within these quotas. A capacity-aware routing algorithm distributes retained pairs among eligible instances of their original experts. Physical compaction then forms smaller communication payloads and variable-size grouped GEMMs, reducing backward traffic, computation, and execution-buffer demand. Experiments on Qwen3-30B-A3B and GLM4.5-106B-A12B show that LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2-44.8% relative to SYMI.

cs.AI↗

arXiv2026

SLT 2026 REAL-TSE Challenge: Real-world Target Speaker Extraction from Conversational Recordings

We introduce the REAL-TSE Challenge, an IEEE SLT 2026 satellite challenge on target speaker extraction~(TSE) from real conversational recordings. Given a multi-speaker mixture and one or more enrollment utterances from a target speaker, participating systems must recover only the target speech. Unlike simulated read-speech benchmarks, REAL-TSE evaluates Mandarin and English recordings that contain natural overlap, reverberation, noise, channel mismatch, and conversational dynamics. The challenge defines two complementary tracks: an Online track for low-latency streaming extraction and an Offline track for full-context processing. Systems are evaluated with Token Error Rate (TER), Speaker Similarity (SpkSim), DNSMOS, and target-speaker activity F1. This overview paper describes the task definition, datasets, baselines, evaluation protocol, submitted systems, condition-wise findings, and lessons for future real-world TSE benchmarks.

eess.AS↗

arXiv2026

Electro-optic comb-driven massive data transmission on monolithically integrated lithium niobate photonic circuits

Driven by large AI models, the demand for datacenter interconnects is outstripping traditional scaling laws. This necessitates a transition toward dense multiplexing techniques, such as wavelength-division multiplexing (WDM), for datacenter environments, transferring long-haul spectral efficiency to short-reach, high-bandwidth applications. Integrated optical frequency combs offer compact, low-cost light sources for future datacenter WDM interconnects. However, existing Kerr comb and mode-locked laser implementations suffer from drawbacks in power efficiency, spectral flatness and phase noise. Their fixed, device-specific channel spacings often conflict with standard WDM grids, and co-integration with high-speed modulators remains challenging. Here, we overcome these limitations by demonstrating high-capacity data transmission using flat-top electro-optic combs on thin-film lithium niobate (TFLN), a platform featuring flat spectra, tunable wavelengths and comb spacing, alongside native compatibility with (de-)multiplexers and high-speed modulators. Using 46 wavelength channels from two lasers and a 50-GHz on-chip electro-optic comb, we realize 16.56 Tb/s coherent transmission over a 98.84-km deployed fiber link between two universities and two datacenters in Hong Kong, verifying its suitability for inter-datacenter connections. To highlight grid reconfigurability, we reconfigure the comb to 100-GHz spacing via an on-chip demultiplexer and demonstrate 10-channel, 80-GBaud coherent transmission at 6.4 Tb/s. We further report a proof-of-concept monolithic WDM transmitter integrating the comb and high-speed modulator, achieving 19-channel, 45-GBaud IM/DD transmission at 1.71 Tb/s. Our work validates TFLN electro-optic combs as flexible, compact and scalable WDM engines for both coherent and IM/DD datacenter links to satisfy rapidly rising bandwidth requirements.

physics.optics↗

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

Efficient Proof Generation for Lean Statements with Multiple Candidate Answers

In formal systems such as Lean 4, efficiently finding answers to mathematical problems and generating verifiable proofs is an important task in automated mathematical reasoning. In the traditional way where an upstream solver provides multiple candidate answers, the common workflow requires searching for proofs for each candidate separately, incurring redundant computational cost. To this end, we propose a joint proof framework that integrates candidate answer selection, answer supplementation, and formal proof generation into a unified reasoning process. At its core is a trained prover that selects given or proposes new answers and generates a proof for at least one answer in a single call. Joint proof can also reuse generated proofs to check for faithfulness errors. For problems with a unique answer, if two distinct answers both obtain verifiable proofs under the same formal statement, this confirms the statement does not faithfully preserve the original problem semantics. On MATH-500, AMC 2023, and AIME 2024/2025/2026, our framework achieves the same formal proof success rate as per-candidate proof generation with about one sixth of the token cost, and identifies at least 6.1% of the formalized statements as unfaithful without requiring additional judges. Our model and code are available at https://github.com/Linppz/Gobble-Prover.

cs.CL↗

arXiv2026

FlexMARL: Graph-Driven Rollout-Training Co-Design for LLM-Based Multi-Agent Reinforcement Learning

Despite algorithm-level innovations for multi-agent reinforcement learning (MARL), the underlying infrastructure for large-scale MARL remains underexplored. Existing training frameworks primarily optimize for single-agent scenarios and fail to capture the graph-structured nature of MARL. This fundamental mismatch leads to three MARL-specific challenges: topology-induced rollout skew, critical path synchronization barriers, and dynamic training demand. To bridge this gap, we propose FlexMARL, a graph-driven rollout-training co-designed framework for large-scale LLM-based MARL. FlexMARL introduces the multi-agent execution graph (MEG) and materializes the intermediate representation as the MEG Store, which provides decentralized multi-table storage and serves as the data hub for system components. Built around the MEG Store, the Rollout Engine develops dependency parallel sampling with hierarchical load balancing, which reduces inference latency and balances intra-agent/inter-agent loads. The Joint Orchestrator exploits an agent-level asynchronous pipeline to mitigate the critical path synchronization barriers while maintaining consistency. The Training Engine achieves on-demand resource allocation by collectively managing training processes for individual agents. The training states of different agents are swapped via location-agnostic communication APIs. Empirical results on a production cluster demonstrate that FlexMARL achieves up to 7.3x speedup and improves hardware utilization by up to 5.6x compared to existing frameworks.

cs.LG↗

arXiv2026

The ALMA Survey of Gas Evolution of PROtoplanetary disks (AGE-PRO): Evolution of dust properties from multi-wavelength observations

Understanding the evolution of solids in protoplanetary discs is important to comprehend the formation of giant planets cores and terrestrial planets. The ALMA Large Program AGE-PRO studied key disc properties in a survey of 30 protoplanetary discs across three evolutionary stages: the embedded phase (0.5-1 Myr), the middle-age phase (1-3 Myr), and the end of the disc lifetime (2-6 Myr). Preliminary dust properties were estimated using 1.3 mm observations, but a multi-wavelength analysis provides a better constraint of important properties such as the dust disc mass, the dust temperature, and the maximum grain size. We present new ALMA 3.1 mm and 1.5 mm (Bands 3 and 5) continuum observations of the complete survey of 30 AGE-PRO discs. We fitted the observed visibilities and derived radial brightness profiles, spatially resolving 20 discs from the AGE-PRO sample at the two studied wavelengths. A multi-wavelength modelling was performed including also the radial profiles from the 1.3 mm observations. Our results replicate trends obtained from the single wavelength analysis: the dust disc size seems to be constant over age and the dust disc mass decreases with time, with a factor $\sim$10-40 uncertainty from opacity assumptions. This result implies much lower gas-to-dust mass ratios for the discs in our sample than previously reported. In addition, the derived spectral indices are consistent with optically thick emission in a large fraction of the AGE-PRO discs, in particular for the Ophiuchus and Lupus discs.

astro-ph.EP↗

arXiv2026

WildMAE: Structure-Aware Masked Autoencoding for Wildlife Image Reconstruction

Camera-trap imagery presents substantial variation in animal appearance, illumination, background clutter, and visual degradation. Masked autoencoders (MAEs) offer a label-free framework for modeling such data through reconstruction of randomly hidden image patches. We present WildMAE, a decoder-side extension of MAE that combines local structure-aware refinement with masked-to-visible cross-attention. A zero-gated convolutional residual path introduces a two-dimensional neighborhood prior, while gated cross-attention allows masked positions to retrieve contextual information from visible-position features without replacing the original Transformer decoder. The associated training recipe progressively introduces normalized-pixel, Sobel-edge, and perceptual objectives. In the original Snapshot Serengeti study, covering 4,744 test images at 75% masking, the reported reconstruction scores were 27.90 versus 24.33 dB PSNR, 0.840 versus 0.780 SSIM, and 0.160 versus 0.320 LPIPS for WildMAE and vanilla MAE, respectively. Additional historical records document favorable average differences under four synthetic degradation conditions and describe component variants. We present the method, experimental observations, and their scope together: the retained metrics are historical protocol-dependent results, rather than a newly reproduced, harmonized comparison or a validation of restoration from genuinely unavailable pixels. WildMAE provides a practical architectural case study for combining global token mixing with spatial and targeted contextual refinement in wildlife masked-image modeling.

cs.CV↗

arXiv2026

Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy

LLM agents that compute the right answer can still answer wrongly. Agent harnesses increasingly keep long tool outputs out of the context window, saving them to files that the agent searches or reads on demand. We ask where this costs accuracy with a real scientific tool: PHREEQC, a widely used geochemistry simulator from the USGS, whose reports reach 846.8k characters and whose determinism lets every printed value be traced. Agents use it to answer 158 execution-verified multiple-choice items. When agents read the simulator's report through a section index rather than in context, GPT-5.1 and GPT-5.2 lose 11.3 and 7.9 points of accuracy, while Claude Opus 4.6 and Claude Sonnet 4.6 show no detectable cost. Re-running all 11,007 logged simulator calls shows that both GPT models compute the keyed value about as often as before; what changes is the evidence they answer from. Under retrieval they often skip the report, answer from their own higher-precision output files, and misround to a neighboring option, a failure we call evidence substitution. With the simulation held fixed, retrieval still costs GPT-5.2 9.4 points. For GPT-5.4, more frequent correct simulations mask an evidence-use loss of about ten points, leaving a net change indistinguishable from zero. Naming the authoritative output raises accuracy by 6.3-7.9 points over plain retrieval for the three GPT models; for GPT-5.1 and GPT-5.2, inlining outputs that fit works as well, while asking for correct rounding does not. Context-management choices in agent harnesses should be judged by where accuracy is lost, not only by final accuracy.

cs.AI↗

arXiv2026

Beyond Compilation: Evaluating Faithful Natural-Language-to-Lean Statement Formalization

Lean verifies that a generated declaration is well typed, but cannot check whether it states what the user intended. We measure how far compilation overstates faithfulness in statement autoformalization without canonical Lean targets. Our criterion requires Lean compilation and agreement of two LLM judges from different model families. On an independent random audit, it agrees with the human majority on 91.5% of cases; it rarely accepts an unfaithful output but rejects some faithful ones. Across eight systems on 227 university-level statements, the gap between compiling and accepted outputs ranges from 1.3 to 29.5 percentage points: our tool-augmented agent compiles 87.2% of the statements, but only 57.7% are accepted. Human review of all the agent's rejections confirms that 23.8 of its 29.5 points are real translation errors. For the gaps of the seven one-shot systems, a two-stage review by Claude Opus 5.5 with Lean and GPT-6.1 Sol, validated on these human labels, finds that 78-100% are real translation errors for six of them. When GPT-5.2, one of the two judges, grades the translations of two models released after it, GPT-6 Astra and Claude Opus 5.5, it accepts 83% of the statements for each, against 21% for one-shot GPT-5.2. For these two models, the same review classifies only 4 of their 54 rejections (7%) as translation errors, so most of their measured gap is judge errors rather than translation errors. This is consistent with these systems erring less while the judge does not: the fewer errors a system makes, the smaller the share of its rejections that are translation errors. Finally, a 2^3 tool ablation of the agent shows that Lean feedback drives most of its compilation gain, and that nearly half of this gain fails the semantic criterion. Code and data are available at https://github.com/grenadeComing/beyond-compilation.

cs.AI↗

arXiv2026

Frame Evidence:A Benchmark for Reference Frames and Visual Evidence in Multi-Image Spatial Reasoning

A spatial answer depends on the reference frame specified in language and its supporting visual evidence. Item accuracy alone obscures whether correctness persists when either changes. We introduce Frame Evidence, a benchmark of 320 matched triplets and 960 Chinese multiple-choice questions from captured and rendered scenes. One family compares three grounded reference frames over fixed images. The other pairs image deletion annotated as insufficient with answer-preserving addition. Joint metrics require correct answers across all three frames and paired evidence variants. Across 20 configurations on 288 held-out questions, GPT three-shot reaches 50.00\% item accuracy, but the best three-frame group accuracy is 5/48 (10.42\%), reached by GPT two-shot. Few-shot prompting and task-specific adaptation improve item accuracy without consistently improving joint correctness. Input controls show where images support matched success, while repeated executions reveal answer instability beneath similar scores.

cs.CV↗

arXiv2026

TS-WAM: Temporal Latent Actions and World-Action Coupling for Vision-Language Robot Control

Many vision-language-action policies rely on action-labeled robot demonstrations and lack an explicit mechanism for predicting future visual states. These limitations hinder learning from action-free videos and can encourage overfitting to demonstration conditions, weakening generalization and robustness. We present TS-WAM, a lightweight Teacher-Student World-Action Model that addresses these challenges through self-supervised temporal learning and explicit future prediction. Its vlm-jepa framework learns from heterogeneous data, including robot demonstrations and action-free human videos, using an EMA target branch without requiring a separately trained external teacher. By coupling vlm-jepa with delta-lwm, TS-WAM learns to predict latent actions from the current observation and instruction and uses them to predict future visual features that guide action generation at inference. A unified end-to-end training architecture jointly optimizes temporal representation learning, visual dynamics, and robot control without a separate latent-action or world-model pretraining stage. The EMA target branch is removed at inference, keeping the deployed policy lightweight. Experiments on LIBERO, LIBERO-Plus, and SimplerEnv, together with real-world manipulation experiments on Franka Research 3 and RevoTron, demonstrate that TS-WAM achieves strong manipulation performance, generalization, and robustness.

cs.RO↗

arXiv2026

BELT : BEhavior-Aligned Low-Rank KV Cache Compression with Learned Top-k Rank Allocation

Long-context inference makes the key-value (KV) cache a major memory bottleneck for large language models. We propose BELT, a low-rank KV cache compression framework. Its key-side objective targets attention distributions rather than raw keys or pre-softmax logits, yielding a quadratic surrogate with a closed-form low-rank solution. A learned global rank allocator uses differentiable soft selection during training and hard selection at deployment to meet a user-specified component budget. Experiments demonstrate improved model quality under aggressive compression and increased decoding throughput at comparable accuracy. Evaluated across three LLMs and multiple benchmarks, BELT achieves up to 70% low-rank KV cache compression while outperforming the strongest baseline by up to 23.3 points on LongBench, and up to 17.8x overall KV cache reduction when combined with quantization. With Triton-based GPU kernels and a quantized cache, BELT delivers up to 2.7x speedup for the attention module and 3.3x end-to-end decoding throughput.

cs.LG↗

arXiv2026

Continuous-wave 148.4 nm Generation in Quasi-phase Matched BaMgF4

We report continuous-wave (CW) vacuum-ultraviolet (VUV) generation at 148.4 nm by second-harmonic generation of 296.8-nm fundamental light in quasi-phase matched (QPM) BaMgF4 (BMF). At the incident power of 170 mW, a VUV output power of approximately 31 pW can be obtained. The quadratic dependence of the signal on the incident power in the QPM region and its attenuation in the ambient environment support the identification of VUV generation. The polarization dependence shows a sine-function shape with a 15% output attenuation from the maximum to the minimum. These results highlight the potential of BMF as a nonlinear medium for CW VUV generation.

physics.optics↗

arXiv2026

Generation of continuous-wave laser light at 148.4 nm using cavity-enhanced second harmonic generation in $BaMgF_4$

The development of a nuclear optical clock based on \textsuperscript{229}Th is expected to influence the next generation of time and frequency standards. At its heart is the need for a continuous wave (CW) at laser at 148.4~nm to drive the clock transition between the ground and excited isomeric state. In this work we experimentally investigate the potential of BaMgF\textsubscript{4} crystals to create a CW solid state laser at the vacuum ultraviolet (VUV) wavelength of 148.4~nm via cavity-enhanced second harmonic generation. For this purpose, a BaMgF\textsubscript{4} crystal was grown, optically polished and periodically poled. The crystal was inserted into a power-enhancement cavity, resonant at the fundamental wavelength of 296.8~nm and the generated laser light at 148.4~nm was characterized. Within this proof-of-concept investigation, a VUV output power of typically ($16\pm1$)~pW is obtained. This marks the first time that this type of crystal is used to generate VUV laser light. The experimental findings are compared to theoretical expectations and provide a clear path for future improvements, which could have implications not only for compact laser sources in the VUV but also solid-state as well as single-ion based \textsuperscript{229}Th nuclear clocks.

physics.optics↗

arXiv2026

CompileMask: Bridging the Gap Between Expressiveness and Efficiency in Attention Masking

Modern attention workloads require diverse masking patterns whose boundaries depend on runtime metadata, yet efficient attention kernels natively support only a small set of patterns. Existing methods for supporting custom masks in these kernels face an expressiveness-efficiency trade-off. Predicate-based methods are expressive but incur substantial preprocessing and execution overhead, whereas interval-based methods execute efficiently but restrict expressiveness through a predefined interval count. We present CompileMask, which addresses this trade-off by connecting expressive mask specifications with efficient attention execution. We design a compiler that analyzes visibility rules during the offline stage and generates a mask-specific preprocess kernel. We further propose MaskSpec, a region-based DSL that lets users express these rules. During the online stage, the generated kernel reads batch metadata and produces a compact Execution Descriptor that lets the attention kernel efficiently check visibility by comparing element indices against precomputed endpoints. Experiments show that CompileMask achieves up to $3.8\times$ speedup over state-of-the-art methods.

cs.DC↗

arXiv2026

The ALMA View of the Edge-on Gomez's Hamburger System: A Highly-Dynamic, Asymmetric Protoplanetary Disk Reveals the Earliest Phases of Giant Planet Formation

Chemical tracers provide some of the strongest observational signatures of ongoing planet formation and localized dynamical perturbations in protoplanetary disks. In particular, sulfur-bearing molecules are predicted to be enhanced in regions of shock heating, ice sublimation, and gravitational instability. Here, we present high-angular-resolution ($\approx$0.$^{\prime\prime}$2) Atacama Large Millimeter/submillimeter Array observations of $^{12}$CO J=3-2, $^{13}$CO J=3-2, CS J=7-6, and SO J$_{\rm N}$=8$_8$-7$_7$ toward the large, edge-on Gomez's Hamburger ('GoHam'; IRAS 18059-3211) disk. We detect a narrow, one-sided arc of SO emission that peaks near a previously-identified gas over-density, suggesting localized heating around an early-stage giant protoplanet or disk fragment. The edge-on geometry of GoHam enables us to place this chemical signature in the broader context of the disk gas and dust structure. To do so, we map the vertical distribution of molecular gas relative to millimeter- and (sub)-micron-sized dust, identify a pronounced north-south continuum asymmetry, and detect non-Keplerian $^{12}$CO and $^{13}$CO emission indicative of a disk wind. We also derive a dynamical stellar mass of 2.2 $\pm$ 0.5 M$_{\odot}$ and a revised dust-extinction-map-based distance of 139 $\pm$ 24 pc, which places GoHam in the outskirts of the Scorpius-Centaurus association. Together, these observations reveal a highly dynamic disk in which localized sulfur chemistry may trace one of the earliest observable stages of wide-separation giant planet formation.

astro-ph.EP↗

arXiv2026

Calibrate the Decisions That Change the Future: On-Policy Post-Training Quantization for Multimodal Large Language Models

Post-training quantization (PTQ) lowers deployment cost for multimodal large language models, but calibration typically reconstructs fixed sequences with local objectives. This overlooks autoregressive feedback: a quantization-induced token change redirects the prefix and changes future states. Yet on-policy coverage alone is insufficient because many decision mismatches barely affect future generation. We propose OnPTQ, an on-policy framework that calibrates on trajectories visited by the current quantized policy. On shared prefixes, OnPTQ identifies quantization-eroded boundaries, evaluates competing tokens through short counterfactual rollouts, and combines current discrepancy with branch consequence into a Decision--Consequence risk. The risk prioritizes critical states, while context anchoring and trajectory refresh preserve multimodal behavior and keep calibration aligned with the updated policy. We further derive a Decision--Consequence bound linking behavioral deviation to current policy discrepancy and action-conditioned future-value span. Across vision--language and omni-modal Qwen models under multiple low-bit settings, OnPTQ improves downstream performance and yields fewer correctness flips against the corresponding Dense/FP16 references, without changing the deployed inference graph.

cs.AI↗