Search arXiv⌕ Search

arXiv subjects

Jia Li

Publications and source records attributed to Jia Li.

At least 73 records · Page 4Linked to original sources
arXiv2026

Envs-FORGE: Frontier-Optimized Reward-Grounded Environment Synthesis for Agent RL

Reinforcement learning (RL) for terminal agents needs executable training environments with reliable rewards and useful difficulty. Fixed recipes such as few-shot, Self-Instruct, and Evol-Instruct apply the same prompting policy to every seed, even when the current policy would benefit from a harder, easier, or simply different task. We present Envs-FORGE, a prompting policy that converts verifier rewards into per-seed environment-synthesis actions. Envs-FORGE estimates seed pass rates, scores six projection--direction actions around a target learning frontier, and solves a per-seed mixed-integer linear program (MILP) to choose the action that conditions generation. The selected action drives synchronized rewriting of the instruction, fixtures, oracle solution, tests, and Docker environment; only gold-verified bundles enter RL training. The indexed MILP form also supports optional soft skill coverage for portfolio planning. On Qwen 3.5 35B, Envs-FORGE improves Pass@1 over Base by 9.2 percentage points on tb-core (40.0% to 49.2%) and 6.4 points on tb-2.0 (23.0% to 29.4%), exceeding the strongest fixed-recipe baseline by 2.4 and 2.1 points. It reaches 77.1% on SWE-bench Verified versus 73.4% for Base, and improves tb-core by 6.8--9.2 points across the evaluated 4B--35B models. All synthesis methods export 100 verified environments and use 2.27M--2.88M synthesis tokens, placing the comparison at the same downstream training-set size and the same operational scale. The source code is available at https://github.com/DataArcTech/DataArc-SynData-Toolkit/.

cs.CL↗
arXiv2026

Finiteness of Equidistant Affine Subspaces

Let \(0\leqslant k<n\), and let \(\mathcal N(k,n)\) denote the supremum of the cardinalities of finite families of affine \(k\)-planes in \(\R^n\) whose pairwise Euclidean separation distances are all equal to one. For \(k=1\), the finiteness of \(\mathcal N(1,n)\) was recently established by Solymosi and Zahl. We extend this result to affine subspaces of arbitrary dimension and prove the fully explicit bound \[ \mathcal N(k,n) \leqslant 2^{\,2^{\,10(k+1)(n-k)}}. \] The main difficulty in higher dimensions is that the rational distance formula degenerates along several possible rank strata of \(\dim(U+V)\), rather than only along the parallel locus. For each fixed rank \(r\), we replace the choice of a nonvanishing Gram minor by a global sum-of-squares expression involving all minors. This produces a polynomial relation \(H_r(x,y)=0\) which encodes both the unit-distance equation and an upper-rank condition, while also vanishing on the diagonal. We propagate this relation by means of the restricted Zariski closure and decompose the resulting algebraic set into an explicitly bounded number of connected Nash submanifolds. An analytic diagonal approximation argument shows that, within each Nash piece, the maximal possible value of \(\dim(U_x+U_y)\) must decrease. Iterating this rank descent over all possible ranks and combining the finitely many standard affine Grassmann charts yields the stated double-exponential upper bound.

math.MG↗
arXiv2026

Sharp Approximation on the Sphere by Linearized Shallow Networks with Analytic Activations

We study the spectral behavior and approximation properties of linearized shallow networks with analytic activation functions on the unit sphere $\mathbb S^d$. Motivated by earlier spherical approximation results for analytic zonal functions \cite{mhaskar1999approximation}, we first derive an asymptotic formula, expressed explicitly in terms of the residues, for the ultraspherical coefficients of a real-analytic function whose relevant singularities are a conjugate pair of simple poles on the imaginary axis. In particular, the formula gives a matching lower bound along the relevant high-degree subsequences and identifies the exponential decay scale and sign oscillation. It applies directly to $\tanh$, the logistic sigmoid, and the rational activations considered here, while the corresponding estimate for $\arctan$ is obtained through differentiation. For such activation functions and quasi-uniform points $\{θ_j^\ast\}_{j=1}^n \subset \mathbb S^d$, functions $f \in \mathcal H^r(\mathbb S^d)$ satisfying the appropriate spectral compatibility condition can be approximated by a linear combination $$f_n(η) = \sum_{j=1}^n a_j σ(θ_j^\ast \cdot η),$$ with the sharp Sobolev rate $$\|f - f_n\|_{\mathcal H^s(\mathbb S^d)} \lesssim n^{-\frac{r-s}{d}} \|f\|_{\mathcal H^r(\mathbb S^d)}.$$ The approximation theorem places the $\mathcal H^s$ norm on the left-hand side and simultaneously provides an explicit normalized $\ell^2$ bound on the coefficients. For spectrally compatible analytic target functions, the same construction further gives the stated analytic convergence rate. Thus, the analysis connects spectral asymptotics and lower bounds at the degrees where the ultraspherical coefficients do not vanish with Sobolev-norm approximation and coefficient control for fixed analytic activation functions.

math.NA↗
arXiv2026

CangjieBench: Benchmarking LLMs on a Low-Resource General-Purpose Programming Language

Large Language Models excel in high-resource programming languages but struggle with low-resource ones. Existing research related to low-resource programming languages primarily focuses on Domain-Specific Languages (DSLs), leaving general-purpose languages that suffer from data scarcity underexplored. To address this gap, we introduce CangjieBench, a contamination-free benchmark for Cangjie, a representative low-resource general-purpose language. The benchmark comprises 248 high-quality samples manually translated from HumanEval and ClassEval, covering both Text-to-Code and Code-to-Code tasks. We conduct a systematic evaluation of diverse LLMs under four settings: Direct Generation, Syntax-Constrained Generation, Retrieval-Augmented Generation (RAG), and Agent. Experiments reveal that Direct Generation performs poorly, whereas Syntax-Constrained Generation offers the best trade-off between accuracy and computational cost. Agent achieve state-of-the-art accuracy but incur high token consumption. Furthermore, we observe that Code-to-Code translation often underperforms Text-to-Code generation, suggesting a negative transfer phenomenon where models overfit to the source language patterns. We hope that our work will offer valuable insights into LLM generalization to unseen and low-resource programming languages. Our code and data are available at https://github.com/cjhCoder7/CangjieBench.

cs.SE↗
arXiv2026

LazyTrain: Limited-resource Allocation toward Zero-waste Yield Optimization in Large Language Model Training

Training large language models on limited hardware is increasingly a scheduling problem across GPU compute, host memory, PCIe transfer, and storage bandwidth. Existing offloading systems reduce GPU residency, and MegaTrain shows that a CPU-master layer-streaming executor can train large models on a single GPU, but fixed checkpointing and placement heuristics still leave communication exposed on the critical path. We propose LazyTrain, an optimization layer over a layer-streaming executor. LazyTrain formulates checkpoint selection, activation placement, recomputation, and CPU-GPU-NVMe communication overlap as a mixed-integer scheduling problem, then executes the solved policy during training. It further couples 8-bit optimizer states with fast gradient clipping as a single Hybrid 8-bit operator: state compression reduces optimizer-state memory, while fast clipping counteracts the additional CPU-side update overhead. Across H800 experiments from Qwen2.5-3B to Qwen3.6-27B, LazyTrain improves sustained TFLOPS over matched baselines runs by approximately 1.24$\times$; RTX 3090 experiments likewise increase the maximum feasible batch size by one at each model scale. In the primary Qwen3.6-27B H800 MetaMathQA run, LazyTrain reaches 219.95 TFLOPS and 1361 tokens/s at batch size 72, peaks at 68.84\,GB of GPU memory, and obtains 95.42\% exact-match accuracy on the full evaluation split. The source code is available at https://github.com/DataArcTech/LazyTrain.

cs.CL↗
arXiv2026

Bayesian-Agent: Posterior-Guided Skill Evolution Across LLM Agent Harnesses

LLM agents increasingly rely on prompts, tools, memory, SOPs, skills, and harness feedback, yet current self-evolution pipelines often update these assets through heuristic reflection or raw success counts. Such updates are brittle when trajectories are sparse, expensive, and context-dependent. We introduce Bayesian-Agent, a native and cross-harness framework that treats reusable agent skills as Bayesian evidence objects. Bayesian-Agent records verified trajectories, maintains posterior beliefs over skill reliability and failure modes, and turns those beliefs into auditable skill actions and model-facing guardrails. This posterior view provides a finite-sample alternative to raw empirical-rate skill updates and frames prompt, context, and harness engineering as inference over the external decision environment. On RealFin-Bench, Bayesian skill evolution matches or improves the raw empirical-rate control and yields large gains on native runs. In the incremental mode, incremental repair improves SOP-Bench from 80\% to 95\%, Lifelong AgentBench from 90\% to 100\%, and RealFin-Bench from 45\% to 65\%. Backend and model-scaling ablations further show that posterior-guided repair can operate across BA native, MiniSWEAgent, and Claude Code, provided the harness produces verifiable task artifacts. The source code is available at https://github.com/DataArcTech/Bayesian-Agent.

cs.CL↗
arXiv2026

Bidirectional Learning of Facial Action Units and Expressions via Structured Semantic Mapping across Heterogeneous Datasets

Facial action unit (AU) detection and facial expression (FE) recognition can be jointly viewed as affective facial behavior tasks, representing fine-grained muscular activations and coarse-grained holistic affective states, respectively. Despite their inherent semantic correlation, existing studies predominantly focus on knowledge transfer from AUs to FEs, while bidirectional learning remains insufficiently explored. In practice, this challenge is further compounded by heterogeneous data conditions, where AU and FE datasets differ in annotation paradigms (frame-level vs.\ clip-level), label granularity, and data availability and diversity, hindering effective joint learning. To address these issues, we propose a Structured Semantic Mapping (SSM) framework for bidirectional AU--FE learning under different data domains and heterogeneous supervision. SSM consists of three key components: (1) a shared visual backbone that learns unified facial representations from dynamic AU and FE videos; (2) semantic mediation via a Textual Semantic Prototype (TSP) module, which constructs structured semantic prototypes from fixed textual descriptions with learnable context prompts for supervision and cross-task alignment in a shared semantic space; and (3) a Dynamic Prior Mapping (DPM) module that incorporates FACS-derived prior knowledge and learns data-adaptive bidirectional association matrices in the textual semantic space for explicit knowledge transfer. Extensive experiments on popular AU detection and FE recognition benchmarks show that SSM consistently outperforms its single-task and multi-task baselines and achieves competitive performance against task-specific methods. The FE-to-AU results further show that holistic expression semantics provides useful supervision for fine-grained AU learning across heterogeneous datasets.

cs.CV↗
arXiv2026

A spectral volume comparison for manifolds with weakly convex boundary

We establish the Bonnet-Myers theorem and the Bishop-Gromov volume comparison theorem in the spectral sense for manifolds with weakly convex boundary. For $n\geq 3$, let $(M^n,g)$ be a simply connected compact smooth $n$-manifold with weakly convex boundary $\partial M$. If there exists a positive function $w\in C^{\infty}(M)$ that satisfies: \begin{equation*} \begin{cases} -\frac{n-1}{n-2}Δw+Λ_{\Ric} w\geq (n-1)w, \enspace in \enspace M, \frac{\partial w}{\partial η}=0, \enspace\enspace\enspace\enspace \enspace\enspace\enspace\enspace\enspace\enspace\enspace\enspace\enspace\enspace\enspace\enspace \enspace\enspace \enspace\enspace on \enspace\partial M, \end{cases} \end{equation*} where $Λ_{\Ric}$ denotes the smallest eigenvalue of the Ricci tensor, $η$ is the unit co-normal vector field of $\partial M$ in $M$, then the diameter of $M$ satisfies $\diam(M)\leq (\frac{\max w}{\min w})^{\frac{n-3}{n-1}}π$.\par If, in addition, $w$ attains its minimum on the boundary $\partial M$, we obtain a sharp upper bound for the volume of $M$: $\Vol(M)\leq \Vol(\bS^n_{+})$, with equality holding if and only if $M^n$ is isometric to the unit round hemisphere $\bS^{n}_{+}$.

math.DG↗
arXiv2026

Kimi K2.5: Visual Agentic Intelligence

We introduce Kimi K2.5, an open-source multimodal agentic model designed to advance general agentic intelligence. K2.5 emphasizes the joint optimization of text and vision so that two modalities enhance each other. This includes a series of techniques such as joint text-vision pre-training, zero-vision SFT, and joint text-vision reinforcement learning. Building on this multimodal foundation, K2.5 introduces Agent Swarm, a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently. Extensive evaluations show that Kimi K2.5 achieves state-of-the-art results across various domains including coding, vision, reasoning, and agentic tasks. Agent Swarm also reduces latency by up to $4.5\times$ over single-agent baselines. We release the post-trained Kimi K2.5 model checkpoint to facilitate future research and real-world applications of agentic intelligence.

cs.CL↗
arXiv2026

Regularized Multitangent Functions and Reduction Theorem

We develop a direct analytic theory of stuffle-regularized multitangent functions and prove their reduction to finite linear combinations of monotangent functions, without using mould calculus. We first establish an asymptotic comparison between one-sided truncated multiple Hurwitz zeta functions and their stuffle regularizations at the natural parameter \(T_N-H(s)\), where \(T_N\) is the harmonic truncation and \(H(s)\) is the harmonic-number function. Applying this comparison to symmetric multitangent truncations yields meromorphic, \(1\)-periodic regularized multitangent functions. An explicit partial-fraction decomposition of each summand, combined with asymptotic estimates for moving truncation ranges, gives formulas for the reduction coefficients in terms of stuffle-regularized multiple zeta values. The constant term and the coefficient of the monotangent \(\mathcal T(1;s)\) are determined from the limits as \(\operatorname{Im}s\to\pm\infty\): both vanish whenever the index contains an entry greater than \(1\), whereas the exceptional indices \(\{1\}^r\) are evaluated through a sine-quotient generating function. As a consequence, we obtain a family of relations among regularized multiple zeta values.

math.NT↗
arXiv2026

Efficient Grammar-Constrained Decoding via Parser Stack Classification

LLMs are widely used to generate structured output like source code or JSON. Grammar-constrained decoding (GCD) can guarantee the syntactic validity of the generated output, by masking out tokens that violate rules specified by a context-free grammar. However, the online computational overhead of existing GCD methods, with latency typically scaling linearly with vocabulary size, limits the throughput of LLMs, especially for models with large vocabularies. To address this issue, we propose PSC, a novel grammar-constrained decoding method. By combining acceptance conditions of all vocabulary tokens into a single classifier of the parser stack during preprocessing, PSC can compute the complete vocabulary mask by checking the parser stack exactly once per decoding step, with time complexity independent of the vocabulary size. Experiments show that PSC computes masks up to 700$\times$ faster than baselines on complex programming language grammars, and up to 30$\times$ faster for schema-conformant JSON; end-to-end LLM throughput with PSC approaches that of unconstrained decoding. We analyze the preprocessing overhead for preprocessing providers and decoding users, and provide a break-even point analysis to help users decide whether to do preprocessing by themselves.

cs.SE↗
arXiv2026

I3DM: Implicit 3D-aware Memory Retrieval and Injection for Consistent Video Scene Generation

Despite remarkable progress in video generation, maintaining long-term scene consistency upon revisiting previously explored areas remains challenging. Existing solutions rely either on explicitly constructing 3D geometry, which suffers from error accumulation and scale ambiguity, or on naive camera Field-of-View (FoV) retrieval, which typically fails under complex occlusions. To overcome these limitations, we propose I3DM, a novel implicit 3D-aware memory mechanism for consistent video scene generation that bypasses explicit 3D reconstruction. At the core of our approach is a 3D-aware memory retrieval strategy, which leverages the intermediate features of a pre-trained Feed-Forward Novel View Synthesis (FF-NVS) model to score view relevance, enabling robust retrieval even in highly occluded scenarios. Furthermore, to fully utilize the retrieved historical frames, we introduce a 3D-aligned memory injection module. This module implicitly warps historical content to the target view and adaptively conditions the generation on reliable warping regions, leading to improved revisit consistency and accurate camera control. Extensive experiments demonstrate that our method outperforms state-of-the-art approaches, achieving superior revisit consistency, generation fidelity, and camera control precision.

cs.CV↗
arXiv2026

To Use AI as Dice of Possibilities with Timing Computation

The dominant noun-based modeling paradigm, grounded in probability theory and committed to pre-specified noun entities as primitive modeling units, is insufficient as a \emph{grammar of thought}: It leaves \emph{timing} outside the computational scope, precluding any adequate representation of the future as an open space of possibilities. This paper addresses three conceptual gaps absent from the existing literature: (1) possibility space -- a framework admitting multiple possible timelines for the same event; (2) timing computation -- the treatment of timing as a computable rather than observed dimension; and (3) causal factum -- the maximal causal efficacy recovered by reasoning backward from possible futures, rather than assumed in advance. Together, these definitions dissolve the confounding problem inherent to noun-based causal inference and provide the foundation for a spontaneously growing causal-reasoning world model. As proof of concept, we instantiate the framework and apply it to longitudinal EHR data from 3,276 breast cancer patients, demonstrating for the first time, to our knowledge, automatic trajectory discovery and counterfactual timing deduction (i.e., a What-If Machine) in a purely data-driven manner.

cs.AI↗
arXiv2026

PerceptionBench: Evaluating Atomic Visual Perception in Multimodal Large Language Models

We introduce PerceptionBench, a benchmark specifically designed to evaluate the atomic visual perception capabilities of Multimodal Large Language Models (MLLMs). Existing benchmarks often fail to isolate perception: holistic evaluations conflate perceptual errors with failures in reasoning or domain knowledge, while application-driven benchmarks only cover narrow, fragmented domains shaped by heuristic designs. To address these limitations, PerceptionBench adopts a bottom-up approach: by diagnosing the earliest failure points in the responses of frontier MLLMs across 42 existing benchmarks, we construct an error taxonomy whose perception branch defines ten atomic perceptual capabilities. Guided by this taxonomy, we construct 3,000 verified questions with short, unambiguous answers, each isolating a single capability, with difficulty stemming from perception rather than reasoning or knowledge. Benchmark results across sixteen frontier MLLMs reveal that atomic perception remains largely unsolved---no model reaches 60\% accuracy, perception-related hallucination is the weakest capability on average, and similar overall scores conceal sharply divergent capability profiles. PerceptionBench thus provides a capability-level standard for measuring and diagnosing the visual perception boundaries of MLLMs.

cs.CV↗
arXiv2026

Tears or Cheers? Benchmarking LLMs via Culturally Elicited Distinct Affective Responses

Culture serves as a fundamental determinant of human affective processing and profoundly shapes how individuals perceive and interpret emotional stimuli. Despite this intrinsic link extant evaluations regarding cultural alignment within Large Language Models primarily prioritize declarative knowledge such as geographical facts or established societal customs. These benchmarks remain insufficient to capture the subjective interpretative variance inherent to diverse sociocultural lenses. To address this limitation, we introduce CEDAR, a multimodal benchmark constructed entirely from scenarios capturing Culturally \underline{\textsc{E}}licited \underline{\textsc{D}}istinct \underline{\textsc{A}}ffective \underline{\textsc{R}}esponses. To construct CEDAR, we implement a novel pipeline that leverages LLM-generated provisional labels to isolate instances yielding cross-cultural emotional distinctions, and subsequently derives reliable ground-truth annotations through rigorous human evaluation. The resulting benchmark comprises 10,962 instances across seven languages and 14 fine-grained emotion categories, with each language including 400 multimodal and 1,166 text-only samples. Comprehensive evaluations of 17 representative multilingual models reveal a dissociation between language consistency and cultural alignment, demonstrating that culturally grounded affective understanding remains a significant challenge for current models.

cs.CL↗
arXiv2026

FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention

Conventional LLMs keep the full KV cache loaded during decoding, causing a severe GPU memory bottleneck for ultra-long context serving. In this report, we propose \textbf{Lookahead Sparse Attention (LSA)}, a novel inference paradigm powered by a Neural Memory Indexer built upon the DeepSeek-V4 architecture. Rather than passively attending to all historical tokens, LSA proactively predicts future context demands and preserves only the query-critical KV chunks in the GPU memory. Crucially, we instantiate this architecture via a \textbf{backbone-free decoupled training} strategy. By formulating the indexer as a standard dual-encoder architecture, we train it independently using standard retrieval training frameworks without ever loading the massive backbone model into GPU memory. We demonstrate that this ``less is more'' paradigm significantly maximizes serving efficiency while acting as an effective attention denoiser in tasks that rely on long-term global memory. Across primary long-context evaluation suites (e.g., LongBench-v2, LongMemEval, and RULER), \texttt{FM-DS-V4} compresses the average physical KV cache footprint down to merely 13.5\% of the full-context baseline, while consistently preserving or slightly elevating downstream accuracy (+0.6\% absolute margin on average). At 1M context, per-decode-token compute drops to 0.30$\times$ of the baseline and GPU KV cache shrinks by 90\% (3.73$\to$0.37 GB), translating into \textbf{2.8$\times$ aggregate throughput and 2.7$\times$ concurrency gains} in PD-disaggregated serving on 8$\times$H20 GPUs.

cs.LG↗
arXiv2026

BrainPilot: Automating Brain Discovery with Agentic Research

Understanding the brain increasingly depends on integrating evidence across scales, modalities, and disciplines. Addressing a single research question therefore requires a coordinated sequence of operations, from surveying prior work to executing analyses and interpreting results in light of domain knowledge. AI agents promise to accelerate this process, but current agents lack domain expertise in brain science, may fabricate claims, drift during multi-step reasoning, and offer few defined points for expert intervention. These failures are especially costly in brain science, where conclusions feed into downstream scientific claims and depend on laboratory-specific expertise and careful human judgment. We present \textbf{BrainPilot} a \textbf{fully open-source} multi-agent system that accelerates brain science research with traceable logs and agent-verified results. A principal investigator (PI) agent coordinates specialist agents grounded in curated domain knowledge: a unified brain science knowledge base containing 7{,}233 indexed items and a skill library of 72 reusable methodology units across seven research domains. Every major step is recorded in the Graph of Trace, an auditable record that links subgoals, tool use, evidence, and claims and allows researchers to follow and inspect the workflow. An Auditor agent further integrates fabrication checking into the workflow. For evaluation, we run three brain science tasks from Agents' Last Exam, introduce our own benchmark, \textbf{BrainPilotBench-v0}, and present additional end-to-end case studies. Across these evaluations, BrainPilot with an open-source backbone model attains performance comparable to state-of-the-art agent framework with less costs.

cs.AI↗