arXiv subjects

Kang Zhou

Publications and source records attributed to Kang Zhou.

At least 19 recordsLinked to original sources

arXiv2026

TRACE: Transition-Aware Residual Control for Multi-Objective Materials Discovery

Multi-objective materials discovery with LLM agents is often limited not only by how many candidates can be proposed, but by how effectively each costly property evaluation informs the next search step. Existing agents mainly store evaluated candidates and their scores, so they know which materials succeeded but not which executable edits caused useful property changes. This makes local refinement difficult when objectives compete and an edit that improves one property may damage another. We propose TRACE, a transition-aware residual control framework that treats evaluated edits as the basic unit of feedback. TRACE records each local refinement as a parent-edit-child transition with observed property deltas, aggregates transition evidence to estimate reusable edit effects, and ranks future edits by their predicted ability to reduce the current candidate's remaining constraint violations while avoiding damage to already satisfied objectives. In a controlled same-backbone comparison, TRACE improves over LLEMA, the state-of-the-art LLM-agent baseline, raising macro-average hit rate from 18.13\% to 25.96\%.

cs.AI↗

arXiv2026

FinEvo-Bench: A Longitudinal Benchmark for Self-Evolving Agents in Professional Financial Workflows

Most agent benchmarks evaluate tasks independently and cannot measure whether experience from one task helps with later tasks. Existing self-evolution benchmarks do not jointly cover professional workflows, open-ended deliverables, and multi-aspect evaluation. We introduce FinEvo-Bench, a longitudinal benchmark with 120 real-case-grounded tasks, 20 business scenes across six financial domains. Institution-provided professional procedures define the required operations and constraints. Eligible institution-provided and publicly documented cases supply the task facts. Each scene contains six related but substantively distinct cases that share a professional procedure and a manually reviewed rubric for task quality and financial compliance. We compare four self-evolving agent scaffolds using the same Qwen3.7-Max backbone and three independently shuffled, globally interleaved task streams. Paired non-evolving controls estimate each scaffold's self-evolution gain from retained experience, while an independent Claude Code scoring agent backed by Claude Opus 4.6 evaluates all outputs. Letta achieves the highest evolved score (91.65) and fewest compliance issues (0.09 per task); Codex achieves the largest self-evolution gain (+19.37). Across scaffolds, the evolving condition raises scores by 9.33-19.37 points and reduces compliance issues by 0.12-0.44 per task. Paired score gains at within-scene ranks 4-6 exceed those at ranks 1-3 by 6.10-8.70 points. In Claude Code, skill-only evolution produces higher task quality and fewer compliance issues than memory-only and combined memory-skill evolution. Across all four scaffolds, rubric feedback also yields higher scores and fewer compliance issues than reference-answer feedback. FinEvo-Bench measures both professional performance and self-evolution ability: how effectively an agent turns prior experience into later improvement.

cs.AI↗

arXiv2026

FinProBench: Evaluating Financial AI Agents with Role-Grounded Rubrics Derived from Professional Deliverables

Evaluating financial AI agents requires criteria aligned with real professional work. Existing rubric methods typically derive criteria from task prompts or model outputs, overlooking tacit standards visible only in practitioner deliverables. We introduce FinProBench, a benchmark for professional financial tasks, and Role-Grounded Rubric Construction (RGRC), a reusable pipeline that derives rubrics from deliverables produced by practitioners in the same role. RGRC comprises four stages: Deliverable Collection, Competency Extraction, Rubric Synthesis, and Validation. Its rubrics capture tacit standards, distinguish quality levels, and transfer across tasks within a role. Before analysis, we classified 57 occupations by deliverable genre into 30 prior-rich conventional roles and 27 prior-sparse role-specialized roles. Across all roles, Prompt-only nearly matches RGRC for conventional roles (89.2% vs. 90.7%), but RGRC substantially outperforms it for role-specialized roles (99.1% vs. 78.0%). This split indicates that prompt engineering can approximate rubrics when conventions are well represented in model priors, while professional grounding is essential for standards beyond those priors. FinProBench is built from 1,723 curated deliverables spanning 57 occupations, 8 financial sub-industries, and 161 deliverable types, and releases an initial evaluation set of 20 complete tasks covering 20 roles in 7 sub-industries. With heterogeneous LLM judges and role-level rubrics, human deliverables rank first on average (73.7 vs. 70.3, 70.2, and 69.6 out of 100), while all four systems show overlapping 95% confidence intervals and complementary strengths. Reusing rubrics at the role level reduces estimated per-task construction effort by 6.7 times relative to authoring each rubric from scratch.

cs.AI↗

arXiv2026

FinPerMA: A Theory-Informed, Event-Grounded Personalized-Memory Benchmark for LLM Agents

Large language model (LLM) agents are increasingly used as personalized assistants in high-stakes domains such as financial advising, yet it remains unclear whether they can maintain and update an individualized user model over long horizons. Existing personalized-memory benchmarks primarily test factual retention or rely on weakly constrained model-generated trajectories, leaving event-driven preference adaptation underexplored. We introduce FinPerMA, an event-grounded benchmark that evaluates personalized memory against frozen longitudinal investor trajectories. Its generation pipeline combines deterministic, theory-informed impact rules, controlled LLM narration, and automated quality screening; a Post-Shock checkpoint isolates whether an agent has integrated a material event into its persistent user model. On 2,994 questions from 276 personas, seven frontier LLMs and up to seven memory configurations remain far from saturated: no full-context configuration exceeds approximately 0.47 overall accuracy or approximately 39% on multiple-choice questions. Attribution analysis shows that summary-based memory often preserves factual details while losing the preference signals needed for personalization; simple retrieval can therefore outperform purpose-built memory systems, with the gap widening after shocks.

cs.AI↗

arXiv2026

Full-Pipeline Inference Optimization for MiMo-V2.5 Series: Pushing Hybrid SWA Efficiency to the Limit

We present a full-pipeline inference optimization for the MiMo-V2.5 model family, which combines Hybrid Sliding Window Attention (Hybrid SWA), sparse Mixture-of-Experts (MoE), and multimodal encoders. While Hybrid SWA can ideally reduce both attention compute and KVCache storage significantly compared to Full Attention, realizing these gains in production requires substantial engineering effort. We systematically optimize the KVCache system with layerwise prefetch, SWA-aware prefix cache trees, and specialized placement strategies, achieving strict $O(W)$ SWA storage and high cache hit rates. We further build GCache, a high-performance distributed cache infrastructure with RDMA-optimized networking, and develop a KVCache-affinity router to reduce computation while preserving load balancing. We also optimize for multimodal inputs, including GPU image preprocessing, parallel video decoding, and multimodal cache sharing. Together, these optimizations constitute the first large-scale LLM serving system in production that efficiently covers the Hybrid SWA + MoE + multimodal composite architecture.

cs.AR↗

arXiv2026

An Empirical Study of Automating Agent Evaluation

Agent evaluation requires assessing complex multi-step behaviors involving tool use and intermediate reasoning, making it costly and expertise-intensive. A natural question arises: can frontier coding assistants reliably automate this evaluation process? Our study shows that simply prompting coding assistants is insufficient for this task. Without domain-specific evaluation knowledge, frontier coding assistants achieve only a 30% execution success rate and produce over-engineered evaluations averaging 12+ metrics per agent, indicating that strong coding ability does not automatically translate to reliable agent evaluation. We introduce EvalAgent, an AI assistant that automates the end-to-end agent evaluation pipeline. EvalAgent encodes evaluation domain expertise as evaluation skills (procedural instructions, reusable code and templates, and dynamically retrieved API documentation) that compose into a trace-based pipeline producing complete evaluation artifacts including metrics, executable code, and reports. To systematically assess generated evaluations, we introduce a meta-evaluation framework alongside AgentEvalBench, a benchmark comprising 20 agents, each paired with evaluation requirements and test scenarios. We further propose the Eval@1 metric to measure whether generated evaluation code both executes and yields meaningful results on the first run. Our experiments show that EvalAgent produces focused evaluations, improving Eval@1 from 17.5% to 65%, and achieving 79.5% human expert preference over baseline approaches. Further ablation studies show that evaluation skills are critical for handling complex evaluation: removing them causes Eval@1 to drop significantly from 65% to 30%.

cs.CL↗

arXiv2026

Universal Interpretation of Hidden Zero and $2$-Split of Tree-Level Amplitudes Using Feynman Diagrams, Part $\mathbf{I}$: ${\rm Tr}(\phi^3)$, NLSM and YM

In this paper, we propose a universal diagrammatic interpretation of hidden zeros and $2$-splits of tree-level amplitudes. Originally developed for ${\rm Tr}(\phi^3)$ amplitudes in our previous work, this interpretation is now extended to tree-level amplitudes in Nonlinear sigma model (NLSM) and Yang-Mills (YM) theories. The interpretation is based on a certain factorization behavior of Feynman diagrams under specific kinematic constraints, which we term shuffle factorization along a specific line (SFASL). This mechanism allows us to separate Feynman diagrams along specific lines after summing over shuffle permutations. When applied to NLSM and YM amplitudes, we perform proper extensions of the SFASL used in the ${\rm Tr}(\phi^3)$ case. Through the SFASL, the interpretation for the hidden zeros and $2$-splits of tree amplitudes of ${\rm Tr}(\phi^3)$, NLSM, and YM can be unified as: the hidden zeros are ascribed to the on-shell condition $k_j^2=0$ of a massless particle, while the $2$-splits are caused by separating each Feynman diagram along two lines, akin to unzipping two zippers.

hep-th↗

arXiv2026

Towards New Hidden Zero and $2$-Split of Loop-Level Feynman Integrands in ${\rm Tr}(\phi^3)$ Model

We extend the hidden zeros and $2$-split of tree-level ${\rm Tr}(\phi^3)$ amplitudes to loop-level Feynman integrands, apart from some physically irrelevant scaleless integrals. Our method is based on a certain factorization mechanism that occurs in Feynman diagrams when summing over shuffle permutations. The loop-level hidden zeros and $2$-split identified in this work differ from those in the literature. In our result, the kinematic conditions for loop-level hidden zeros and $2$-split are remarkably simple. Their connection is as tight as at tree-level, with the same procedure for obtaining the $2$-split condition from the zero condition. The resulting $2$-split formula at loop-level represents a generalization of that at tree-level: the $L$-loop integrand is expressed as a sum over $L+1$ terms, each of which exhibits a $2$-split structure.

hep-th↗

arXiv2026

Can Locality, Unitarity, and Hidden Zeros Completely Determine Tree-Level Amplitudes?

In this note, we address the question of whether locality, unitarity, and newly discovered hidden zeros can completely determine tree-level amplitudes, from the perspective of soft limit. We reconstruct the single-soft theorems of tree YM amplitudes and the double-soft theorems of tree NLSM amplitudes from locality, unitarity, and hidden zeros. A series of studies have shown that the full YM and NLSM amplitudes can be constructed from these soft theorems; therefore, we conclude that locality, unitarity, and hidden zeros completely determine the tree-level YM and NLSM amplitudes.

hep-th↗

arXiv2026

UltrasoundAgents: Hierarchical Multi-Agent Evidence-Chain Reasoning for Breast Ultrasound Diagnosis

Breast ultrasound diagnosis typically proceeds from global lesion localization to local sign assessment and then evidence integration to assign a BI-RADS category and determine benignity or malignancy. Many existing methods rely on end-to-end prediction or provide only weakly grounded evidence, which can miss fine-grained lesion cues and limit auditability and clinical review. To align with the clinical workflow and improve evidence traceability, we propose a hierarchical multi-agent framework, termed UltrasoundAgents. A main agent localizes the lesion in the full image and triggers a crop-and-zoom operation. A sub-agent analyzes the local view and predicts four clinically relevant attributes, namely echogenicity pattern, calcification, boundary type, and edge (margin) morphology. The main agent then integrates these structured attributes to perform evidence-based reasoning and output the BI-RADS category and the malignancy prediction, while producing reviewable intermediate evidence. Furthermore, hierarchical multi-agent training often suffers from error propagation, difficult credit assignment, and sparse rewards. To alleviate this and improve training stability, we introduce a decoupled progressive training strategy. We first train the attribute agent, then train the main agent with oracle attributes to learn robust attribute-based reasoning, and finally apply corrective trajectory self-distillation with spatial supervision to build high-quality trajectories for supervised fine-tuning, yielding a deployable end-to-end policy. Experiments show consistent gains over strong vision-language baselines in diagnostic accuracy and attribute agreement, together with structured evidence and traceable reasoning.

cs.CV↗

arXiv2026

Learning to Ideate for Machine Learning Engineering Agents

Existing machine learning engineering (MLE) agents struggle to iteratively optimize their implemented algorithms for effectiveness. To address this, we introduce MLE-Ideator, a dual-agent framework that separates ideation from implementation. In our system, an implementation agent can request strategic help from a dedicated Ideator. We show this approach is effective in two ways. First, in a training-free setup, our framework significantly outperforms implementation-only agent baselines on MLE-Bench. Second, we demonstrate that the Ideator can be trained with reinforcement learning (RL) to generate more effective ideas. With only 1K training samples from 10 MLE tasks, our RL-trained Qwen3-8B Ideator achieves an 11.5% relative improvement compared to its untrained counterpart and surpasses Claude Sonnet 3.5. These results highlights a promising path toward training strategic AI systems for scientific discovery.

cs.CL↗

arXiv2026

MiMo-V2-Flash Technical Report

We present MiMo-V2-Flash, a Mixture-of-Experts (MoE) model with 309B total parameters and 15B active parameters, designed for fast, strong reasoning and agentic capabilities. MiMo-V2-Flash adopts a hybrid attention architecture that interleaves Sliding Window Attention (SWA) with global attention, with a 128-token sliding window under a 5:1 hybrid ratio. The model is pre-trained on 27 trillion tokens with Multi-Token Prediction (MTP), employing a native 32k context length and subsequently extended to 256k. To efficiently scale post-training compute, MiMo-V2-Flash introduces a novel Multi-Teacher On-Policy Distillation (MOPD) paradigm. In this framework, domain-specialized teachers (e.g., trained via large-scale reinforcement learning) provide dense and token-level reward, enabling the student model to perfectly master teacher expertise. MiMo-V2-Flash rivals top-tier open-weight models such as DeepSeek-V3.2 and Kimi-K2, despite using only 1/2 and 1/3 of their total parameters, respectively. During inference, by repurposing MTP as a draft model for speculative decoding, MiMo-V2-Flash achieves up to 3.6 acceptance length and 2.6x decoding speedup with three MTP layers. We open-source both the model weights and the three-layer MTP weights to foster open research and community collaboration.

cs.CL↗

arXiv2026

Generalized $2$-split for higher-derivative YM and GR amplitudes at tree-level

We study the generalized $2$-split of higher-derivative amplitudes, including Yang-Mills (YM) and Gravity (GR) amplitudes with special insertions of higher-derivative vertices, by expanding them into ${\rm YM}\oplus{\rm BAS}$, ${\rm GR}\oplus{\rm YM}$, and ${\rm GR}\oplus{\rm YM}\oplus{\rm BAS}$ amplitude, respectively. By leveraging the established $2$-split properties of these constituent theories, we show that these higher-derivative amplitudes -- which also exhibit another newly discovered phenomenon called hidden zero -- do not factorize into a single product of two currents. Instead, their factorization universally appears as a sum of multiple $2$-split contributions.

hep-th↗

arXiv2025

MiMo-Audio: Audio Language Models are Few-Shot Learners

Existing audio language models typically rely on task-specific fine-tuning to accomplish particular audio tasks. In contrast, humans are able to generalize to new audio tasks with only a few examples or simple instructions. GPT-3 has shown that scaling next-token prediction pretraining enables strong generalization capabilities in text, and we believe this paradigm is equally applicable to the audio domain. By scaling MiMo-Audio's pretraining data to over one hundred million of hours, we observe the emergence of few-shot learning capabilities across a diverse set of audio tasks. We develop a systematic evaluation of these capabilities and find that MiMo-Audio-7B-Base achieves SOTA performance on both speech intelligence and audio understanding benchmarks among open-source models. Beyond standard metrics, MiMo-Audio-7B-Base generalizes to tasks absent from its training data, such as voice conversion, style transfer, and speech editing. MiMo-Audio-7B-Base also demonstrates powerful speech continuation capabilities, capable of generating highly realistic talk shows, recitations, livestreaming and debates. At the post-training stage, we curate a diverse instruction-tuning corpus and introduce thinking mechanisms into both audio understanding and generation. MiMo-Audio-7B-Instruct achieves open-source SOTA on audio understanding benchmarks (MMSU, MMAU, MMAR, MMAU-Pro), spoken dialogue benchmarks (Big Bench Audio, MultiChallenge Audio) and instruct-TTS evaluations, approaching or surpassing closed-source models. Model checkpoints and full evaluation suite are available at https://github.com/XiaomiMiMo/MiMo-Audio.

cs.CL↗

arXiv2025

Diffusion Language Model Inference with Monte Carlo Tree Search

Diffusion language models (DLMs) have recently emerged as a compelling alternative to autoregressive generation, offering parallel generation and improved global coherence. During inference, DLMs generate text by iteratively denoising masked sequences in parallel; however, determining which positions to unmask and which tokens to commit forms a large combinatorial search problem. Existing inference methods approximate this search using heuristics, which often yield suboptimal decoding paths; other approaches instead rely on additional training to guide token selection. To introduce a principled search mechanism for DLMs inference, we introduce MEDAL, an inference-time scaling framework that integrates Monte Carlo Tree SEarch initialization for Diffusion LAnguage Model inference. We employ Monte Carlo Tree Search at the initialization stage to explore promising unmasking trajectories, providing a robust starting point for subsequent refinement. This design enables efficient inference-time scaling, allowing generation quality to improve as the search budget increases, without additional training. Across multiple benchmarks, MEDAL achieves up to 22.0% improvement over existing inference strategies, establishing a new paradigm for search-based inference in DLMs.

cs.CL↗

arXiv2025

Beyond Correctness: Rewarding Faithful Reasoning in Retrieval-Augmented Generation

Inspired by the success of reinforcement learning (RL) in Large Language Model (LLM) training for domains like math and code, recent work has begun training LLMs to dynamically plan, query, and reason with search engines as tools -- a paradigm increasingly referred to as agentic search. Although these methods achieve performance improvement across popular short-form QA benchmarks, many prioritize final answer correctness while overlooking the quality of intermediate reasoning steps, which may lead to chain-of-thought unfaithfulness. In this paper, we first introduce a comprehensive evaluation framework for agentic search, covering three distinct faithfulness metrics: Think-Search faithfulness, Information-Think faithfulness, and Think-Answer faithfulness. Our evaluations reveal that canonical agentic search systems trained through Reinforcement Learning from Verifiable Reward (RLVR) using episode-level outcome-based reward -- including Search-R1 and ReSearch -- have significant room for improvement on these faithfulness dimensions. To foster faithful reasoning in agentic search, we introduce VERITAS (Verifying Entailed Reasoning through Intermediate Traceability in Agentic Search), a novel framework that integrates fine-grained turn-level faithfulness rewards into the reinforcement learning process. Our experiments show that models trained with \ours not only significantly improve reasoning faithfulness, but also achieve better task performance compared to baselines trained against episode-level outcome-based reward.

cs.CL↗

arXiv2025

Hidden zeros for higher-derivative YM and GR amplitudes at tree-level

We extend the recently discovered phenomenon of hidden zeros to tree amplitudes for Yang-Mills (YM) and general relativity (GR) theories with higher-derivative interactions. This includes gluon amplitudes with a single insertion of the local $F^3$ operator, as well as graviton amplitudes at sub-leading and sub-sub-leading orders in the low-energy expansion of bosonic closed string amplitudes -- referred to as $R^2$ and $R^3$ amplitudes, respectively. The kinematic condition for hidden zeros leads to unavoidable propagator singularities in unordered graviton amplitudes. We investigate in detail the systematic cancellation of these divergences, which resolves ambiguities in the proof of hidden zeros. Our approach is based on universal expansions that express tree amplitudes as linear combinations of bi-adjoint scalar amplitudes.

hep-th↗

arXiv2025

IPR: Intelligent Prompt Routing with User-Controlled Quality-Cost Trade-offs

Routing incoming queries to the most cost-effective LLM while maintaining response quality poses a fundamental challenge in optimizing performance-cost trade-offs for large-scale commercial systems. We present IPR\, -- \,a quality-constrained \textbf{I}ntelligent \textbf{P}rompt \textbf{R}outing framework that dynamically selects optimal models based on predicted response quality and user-specified tolerance levels. IPR introduces three key innovations: (1) a modular architecture with lightweight quality estimators trained on 1.5M prompts annotated with calibrated quality scores, enabling fine-grained quality prediction across model families; (2) a user-controlled routing mechanism with tolerance parameter $\tau \in [0,1]$ that provides explicit control over quality-cost trade-offs; and (3) an extensible design using frozen encoders with model-specific adapters, reducing new model integration from days to hours. To rigorously train and evaluate IPR, we curate an industrial-level dataset IPRBench\footnote{IPRBench will be released upon legal approval.}, a comprehensive benchmark containing 1.5 million examples with response quality annotations across 11 LLM candidates. Deployed on a major cloud platform, IPR achieves 43.9\% cost reduction while maintaining quality parity with the strongest model in the Claude family and processes requests with sub-150ms latency. The deployed system and additional product details are publicly available at https://aws.amazon.com/bedrock/intelligent-prompt-routing/

cs.LG↗