Searcharxiv⌕ Search

arXiv subjects

Jun Sun

Publications and source records attributed to Jun Sun.

At least 19 recordsLinked to original sources
arXiv2026

Beyond Gain: Measuring Skill Internalization for Agent Skill Retirement

LLM agents retrieve skills from public hubs that hold up to about 100,000 entries and serve many models. As models improve, a hub must find the skills a given model has internalized, meaning it can do without them, and retire them for that model, yet internalization has not been measured with known reliability. Published rules that compare runs with and without a skill retire on the gain alone. The gain is near zero both for an internalized skill and when neither the model nor the skill delivers the capability, a state we call uncovered. We instead withhold the skill, score the model on a task synthesized from it, and trust a high unaided score only when another model gains from the skill, since nearly half of high scores come from tasks no model benefits from. Under four models, a published gain rule applied with one task per skill removes over half of the uncovered skills. On SkillsBench, 28 of its 31 retirements rest on probes no run passed in full, while retiring on the unaided score costs no detectable reward. Unaided verdicts reproduce across tasks more often than the gain rule's (Cohen's kappa 0.42 against 0.23) but differ across models, so a hub must read them per model. We release 1,251 skills with tasks and item scores under four models.

cs.AI↗
arXiv2026

SkillRep: Characterizing and Understanding the Representation Design Space of Agent Skill Retrieval

Retrieving the right skill from a large library is essential for LLM agents, yet requests and skills are written for different purposes and in different words. Existing retrievers narrow this request-skill gap with one or two representation choices each, so it is unknown whether their gains hold in combination. No prior work measures the gap itself, so it is also unclear which part of it each choice narrows. We define a representation design space by five choices -- encoder fine-tuning, query-side and document-side expansion, background projection and a lexical channel -- and evaluate them together for the first time, as modules of one retriever, SkillRep. With a 0.6B encoder and LLM expansion, SkillRep improves MRR@10 on SRA-Bench over the strongest same-size retriever by 7.02 points with the official skill pool and 8.79 with UniPool, a 111,247-skill pool we build from public registries, and is best on SkillsBench. The value of each choice depends on the others. Fine-tuning, for instance, reverses the ranking of the four training-free modules and turns them from complementary into overlapping. To understand why, we measure the gap with three quantities: background, the embedding component that does not separate skills; unsaid vocabulary, the gold skill's content a request leaves unsaid; and dilution, the information lost when a skill is encoded into a single vector. They show which part each module narrows and by what mechanism. This reveals that modules narrowing the same part compete for it, while those narrowing different parts complement each other.

cs.IR↗
arXiv2026

PAFS: Perturbation-Aware Firing Scaffolds for Robust Event-Based SNNs

Event-based spiking neural networks (SNNs) encode information through temporal firing dynamics, but the same stateful computation makes their predictions sensitive to event loss, timestamp jitter, and speed variation. These perturbations alter evidence accumulation, yet late-layer firing trajectories can retain error-diagnostic structure that is absent from the final logits alone. We introduce Perturbation-Aware Firing Scaffolds (PAFS), a labeled post-training calibration method that exploits this structure without updating the pretrained backbone. PAFS associates standardized late-layer leaky integrate-and-fire trajectories with the class-probability residuals of an Adapter anchor, then applies singular value decomposition to extract compact residual-aligned channel directions. At inference, firing trajectories from the same backbone pass are projected onto this fixed low-rank scaffold and combined with anchor statistics to predict a residual logit correction. Across four matched dataset-backbone settings spanning three event datasets and two spiking Transformer architectures, PAFS improves corrupted-average and worst-condition accuracy over the corresponding Adapter anchors, with maximum gains of 3.06 and 3.65 percentage points. In the DailyAction-DVS/Spikingformer setting, the added trainable parameters amount to approximately 0.34% of the frozen backbone.

cs.NE↗
arXiv2026

WenSight-Decision-1 for Direct Decisions under E-commerce Rules

E-commerce decisions require applying natural-language rules to the available case evidence. We introduce WenSight-Decision-1, a shared model that scores supplied options directly from rules and case information for six operational tasks, with examples spanning six languages. We fine-tune Qwen3.5-9B with reference options as supervision, using the same option scores for training and inference. On 1,440 synthetic operational cases reserved for final evaluation, accuracy rises from 61.60% for Qwen3.5-9B to 92.92% for WenSight-Decision-1 under identical inputs and scoring. Joint fine-tuning improves all six tasks, bringing heterogeneous operational judgments under a common decision interface.

cs.CL↗
arXiv2026

Mut4All: Fuzzing Compilers via LLM-Synthesized Mutators Learned from Bug Reports

Mutation-based fuzzing is effective for uncovering compiler bugs, but designing high-quality mutators for modern languages with complex constructs (e.g., templates, macros) remains challenging. Existing methods rely heavily on manual design or human-in-the-loop correction, limiting scalability and cross-language generalizability. We present Mut4All, a fully automated, language-agnostic framework that synthesizes mutators using Large Language Models (LLMs) and compiler-specific knowledge from bug reports. It consists of three agents: (1) a mutator invention agent that identifies mutation targets and generates mutator metadata using compiler-related insights; (2) a mutator implementation synthesis agent, fine-tuned to produce initial implementations; and (3) a mutator refinement agent that verifies and corrects the mutators via unit-test feedback. Mut4All processes 1400 bug reports (700 Rust, 700 C++), yielding 444 Rust and 561 C++ mutators at ~$0.08 each via GPT-4o. Our customized fuzzer, using these mutators, finds 62 bugs in Rust compilers (44 new, 32 fixed) and 38 bugs in C++ compilers (17 new, 3 fixed). Mut4All outperforms existing methods in both unique crash detection and coverage, ranking first on Rust and second on C++.

cs.SE↗
arXiv2026

Representation Transitions Reveal Emerging Safety Risks in Multi-Turn LLM Agents

Multi-turn attacks on agentic systems can compose individually permissible actions into harmful outcomes, challenging defenses that assess actions or states in isolation. We show that such attacks leave a detectable signature in the agent's internal representations: harmful behavior emerges as an accumulated representation transition across context updates, whose triggering context can be identified from the same signal. We further find that naive aggregation is confounded by benign representation drift, as a contrastive safety direction need not assign zero to benign transitions. We address this by denoising the direction, anchoring benign traffic at zero and removing its leading variation directions, with no runtime cost. These findings motivate DART, a runtime framework that detects and attributes representation shifts and intervenes with targeted reminders. Across six models and two multi-turn benchmarks, DART reduces attack success from 84% to 25% on MT-AgentRisk, catching every attack at a mean false-alarm rate of 12%, and from 97% to 52% on ASEval, at costs in benign non-refusal of 8% and 0%, respectively. On MT-AgentRisk, it outperforms ToolShield, the state-of-the-art multi-turn defense, on all six models: under the same protocol, ToolShield reaches only 55%. Denoising is critical: on ASEval, the undenoised monitor catches only 7%-40% of attacks, while the denoised monitor catches 60%-85%. The same monitor covers single-turn indirect injection without modification and adds only 0.14-0.56 s overhead per monitored step without requiring an auxiliary model, making it a lightweight complement to computation-heavy speculative defenses.

cs.LG↗
arXiv2026

Beyond the Matrix Sign: Quadratic Spectral Descent

Muon emerges as a strong competitor of the AdamW for LLM pretraining, because the matrix-wise update it employs can potentially incur smaller second-order penalty than the once dominating AdamW, which performs coordinate-wise update. However, the spectral flattening procedure in Muon is quite debatable since it discards the spectral amplitude information totally. To seek for better spectral allocation (and the associated spectral subspace), we propose to solve the quadratic model of loss function under the spectral norm constraint \textit{directly} (i.e., in a genuinely Newtonian way) and thus obtaining the Quadratic Spectral Descent (QSD) algorithm. In contrast, many existing curvature-aware methods either exploit the second-order information in an \textit{implicit} way by changing the weight update geometry (such as Mousse, FISMO) or rely on strong assumptions (such as the weight displacement isotropy assumption in Newton-Muon). QSD's potential advantage over these methods is best illustrated in the isotropic curvature scenario, where Mousse, FISMO and Newton-Muon all reduce to Muon while the spectral allocation in QSD is still \textit{non-flat} (since the spectral allocation in QSD depends on the \textit{gradient to curvature ratio}). Meanwhile, to control the complexity of QSD, we employ inversion-free K-FAC and \textit{online} Frank-Wolfe update which is essentially a matrix sign operator. Overall, the complexity increase can be rather mild. Experiments on GPT pre-training show that QSD consistently improves validation loss over Muon and recent Muon variants, while achieving up to an $8.49\%$ wall-clock speedup at matched validation loss.

cs.LG↗
arXiv2026

Which Influence Are We Estimating? The Role of Counterfactual Specifications in Data Attribution

Estimating the influence of training examples on model behavior is essential for data debugging, valuation, and attribution. Existing influence estimators often produce incompatible rankings, which are commonly ascribed to approximation error. We argue that a more fundamental source of disagreement is specification mismatch: influence depends on the behavior being attributed, the intervention applied to each training example, and the counterfactual training process that maps the intervention to a model response. These choices are especially important when the target behavior requires a tractable surrogate, such as query loss, a logit, or a margin. We formalize influence as a counterfactual estimand, distinguish specification mismatch across estimands from approximation error in estimating a fixed estimand, and organize representative estimators by their implied specifications. We further derive a local decomposition that exposes how behavior signals, training signals, and counterfactual parameter responses interact. Controlled experiments show that exact estimands under different specifications can induce different rankings, whereas approximation error grows as perturbations move farther from their linearization points. Experiments on noisy label detection and LLM attribution show that specification choices significantly affect attribution quality, especially for the choice of behavior surrogate. Behavior-aligned specifications can identify target-specific training examples obscured by default loss-based or similarity-based specifications. These results establish specification analysis as a necessary first step for interpreting and comparing data influence estimators.

cs.AI↗
arXiv2026

Hamilton-Type Gradient Estimates for Porous Medium and Fast Diffusion Equations on Riemannian Manifolds for all exponents

We establish local Hamilton-type gradient estimates for positive $C^{2,1}$ solutions of $u_t=Δu^m$ on complete Riemannian manifolds whose Ricci curvature is bounded from below. For every fixed $m>1$ and every fixed $0 0$ and $C=C(m,n)>0$ such that a solution $0<u\leqslant A$ in $B_{2R}(x_0)\times(t_0-T,t_0]$ satisfies the following local gradient estimate \begin{equation*} \sup_{B_R(x_0)\times(t_0-T/2,t_0]}|\nabla u^β| \leqslant C A^β\left(\frac1R+\sqrt{k}+\frac{A^{(1-m)/2}}{\sqrt T}\right), \end{equation*} where $\mathrm{Ric}_M\geqslant-k$. The proof uses an intrinsic quantitative alternative to locate, around each prescribed positive point, a cylinder on which the solution has a controlled upper-to-lower ratio. A local gradient estimate on this cylinder is combined with a stopping argument. As a consequence, every uniformly bounded positive ancient solution on a connected complete manifold with nonnegative Ricci curvature is constant.

math.DG↗
arXiv2026

Hamilton's gradient estimates and Liouville theorems for $u_{t}=Δu^{m}+au\log u+bu$ on Riemannian manifolds

In this paper, we apply Nash-Moser iteration and Saloff-Coste's Sobolev inequalities to derive gradient estimates for positive solutions to a class of nonlinear parabolic equations of the form \begin{equation*} u_{t}=Δu^{m}+a(x, t)u\log u+b(x, t)u, \quad 1-\frac{2}{n} 1$ or $m<1$, respectively. As consequences, we obtain Liouville type theorems for the corresponding constant coefficient equations when \begin{equation*} 1-\frac{1}{\sqrt{n-1}}<m<1\quad \text{and}\quad 1<m<1+\frac{1}{\sqrt{n-1}}. \end{equation*}

math.AP↗
arXiv2026

Metavalent Bonding-Induced Phonon Hardening and Giant Anharmonicity in BeO

The search for materials with intrinsically low thermal conductivity ($κ_L$) is critical for energy applications, yet conventional descriptors often fail to capture the complex interplay between bonding and lattice dynamics. Here, first-principles calculations are used to contrast the thermal transport in covalent zincblende (zb) and metavalent rocksalt (rs) BeO. We find that the metavalent bonding in rs-BeO enhances lattice anharmonicity, activating multi-phonon scattering channels and suppressing phonon transport. This results in an ultralow $κ_L$ of 24 W m$^{-1}$ K$^{-1}$ at 300 K, starkly contrasting with the zb phase (357 W m$^{-1}$ K$^{-1}$). Accurately modeling such strongly anharmonic systems requires explicit inclusion of temperature-dependent phonon renormalization and four-phonon scattering. These contributions, negligible in zb-BeO, are essential for high-precision calculations of the severely suppressed $κ_L$ in rs-BeO. Finally, we identify three key indicators to guide the discovery of metavalently bonded, incipient-metallic materials: (i) an NaCl-type crystal structure, (ii) large Grüneisen parameters ($\textgreater$2), and (iii) a breakdown of the Lyddane-Sachs-Teller relation. These findings provide microscopic insight into thermal transport suppression by metavalent bonding and offer a predictive framework for identifying promising thermoelectrics and phase-change materials.

cond-mat.mtrl-sci↗
arXiv2026

Soft Posterior Speaker Injection for Multi-Talker Speech Recognition

Multi-talker automatic speech recognition (MT-ASR) remains challenging in the presence of overlapping speech. Hard segmentation introduces irreversible errors, whereas serialized output training (SOT) avoids explicit segmentation but does not condition a pretrained encoder on speaker activity. We propose Soft Posterior Speaker Injection (SPSI). A Soft Posterior Head predicts per-frame speaker posteriors $\hat{\mathbf{P}}$ and injects them into Whisper through Multi-layer Feature-wise Linear Modulation (MFLM) and Speaker Memory Prompts (SMP). The benefit of SPSI is largest where overlap is heaviest and under domain transfer. On controlled two-speaker LibriSpeech overlap, SPSI reduces concatenated minimum-permutation word error rate (cpWER) from $61.5\%$ to $60.0\%$ in the high-overlap bin, and from $51.9\%$ to $51.0\%$ on the full set, relative to SOT. By contrast, Speaker CE, SD-CTC, SA-DiCoW, and Pipeline (oracle/est.\ VAD) do not outperform SOT. Freeze-posterior overlap-heavy adaptation reduces held-out LibriCSS cpWER from $42.3\%$ to $36.8\%$ on sessions $8$--$9$, a $5.5$-point gain over SOT. The source code is available at https://github.com/HackerHyper/SPSI.

cs.SD↗
arXiv2026

Digital Self-Interference Cancellation in Full-Duplex Radios: A Fundamental Limit Perspective

D-SIC is of crucial importance for the implementation of IBFD radios. Unfortunately, the achievable performance limit remains underexplored. To fill this gap, in this paper we aim to explore the performance limit, i.e., the minimum residual self-interference (RSI) of the most commonly used PH canceller, and provide the achievable pilot design accordingly. To this end, we first conduct a systematic analysis of the RSI power for the PH canceller, which takes into account both the truncation-induced error and the noise-induced error, whereas the former is usually ignored in the existing works. To simplify the performance analysis of RSI power, we employ the generalized Laguerre polynomial (GLP)-based PH canceller instead of the conventional monomial-based one, due to the appealing orthogonality property of the GLP for Gaussian inputs. With the GLP representation of the PH canceller, we further prove that the least-squares channel estimator is asymptotically unbiased, thus demonstrating the asymptotic optimality of Gaussian pilot sequences. Moreover, for the pilot sequence with a finite length, a succinct criterion for minimizing the RSI, namely, the condition-number-to-minimum eigenvalue ratio (CMER) criterion, which essentially balances the truncation-induced and noise-induced error, is presented. By contrast, the existing works normally consider the latter only. Interestingly, it is revealed that an appropriate PAPR of the pilot sequence is of critical importance to achieve the above balance. Simulation results demonstrate that the pilot sequence optimized according to our proposed CMER criterion can achieve an RSI as low as -87.3 dBm, which is over 14 dB lower than that of HE-LTF and over 6 dB lower than that of the state-of-the-art pilot sequence proposed in [1], provided that the order of the PH canceller is no higher than 9 because of the complexity constraint.

eess.SP↗
arXiv2026

Four-phonon scattering and coherent heat transport in ultrawide-bandgap SrSnO3

SrSnO3 is a promising ultrawide-bandgap perovskite oxide whose thermal transport is governed by structural distortions and anharmonic lattice dynamics. Here, we investigate the lattice thermal conductivity (kL) of orthorhombic and cubic SrSnO3 within a unified first-principles framework combining self-consistent phonon renormalization, three- and four-phonon scattering, and coherent heat transport. Bonding analysis reveals a rigid Sn-O octahedral framework embedded in a weakly bonded Sr sublattice, giving rise to low-frequency vibrational modes susceptible to strong anharmonic effects. Four-phonon scattering is identified as a key mechanism limiting particle-like heat conduction, reducing the Peierls thermal conductivity by 19.4% at 300 K in the orthorhombic phase and by 52.1% at 1300 K in the cubic phase, while the coherent contribution provides a finite channel that partially compensates this reduction. We further show that the apparent agreement between three-phonon calculations and experimental thermal conductivity at room temperature is not indicative of a complete physical description. Instead, it arises from a near cancellation between four-phonon suppression of the particle-like channel and the neglected coherent contribution. This cancellation breaks down when the full temperature dependence is considered, where only the combined treatment improves agreement with the experimentally observed scaling behavior. A physically consistent description of kL therefore requires phonon renormalization, four-phonon scattering, and coherent transport to be treated on equal footing rather than inferred from three-phonon agreement at a single temperature. These results provide microscopic insight into thermal transport in ultrawide-bandgap stannate perovskites and establish a benchmark for anharmonic transport in strongly distorted oxides.

cond-mat.mtrl-sci↗
arXiv2026

SkillCorpus: Consolidating and Evaluating the Open Skill Ecosystem for Real-World LLM Agents

Agent skills, SKILL files that package reusable procedural knowledge for an LLM agent, are a popular mechanism for extending agent capabilities. Public repositories now host them in large and growing numbers, yet these artifacts are fragmented, redundant, and uneven in quality, and their value in practice is unclear. A core question remains open, namely how to consolidate this open-source SKILL ecosystem into a single usable corpus, and what bounds its benefit on real-world agent tasks. We present SkillCorpus, a framework that aggregates, curates, matches, and evaluates the open skill ecosystem at scale. It filters ~821,000 crawled skills through a multi-stage pipeline into 96,401 skills organised by a 16-class taxonomy and three quality facets (utility, robustness, safety), and pairs them with a fine-tuned retrieval-and-selection stack that matches task-relevant skills. We evaluate end-to-end across three benchmarks (SkillsBench, GDPVal, QwenClawBench), two harnesses, and two open backbones with a frontier robustness check. Integrating SkillCorpus yields consistent gains across all three benchmarks, largest on SkillsBench (+7.5 pp). An operational analysis traces the gains to a coverage boundary and a harness boundary. SkillCorpus is, to our knowledge, the first end-to-end account of when a curated, retrieval-served community corpus improves real agent tasks, and where it does not. The dataset, models, and code are available at https://github.com/EverMind-AI/SkillCorpus.

cs.CL↗
arXiv2026

Not All Tokens Are Equal: Region-Aware Consistency Repair of Backdoors in MLLMs

MLLMs are increasingly deployed in user-facing applications, yet they inherit backdoor risks from the pipelines used to construct them: triggers may reside in images, texts, or both. Existing model-level backdoor removal methods, largely designed for conventional classifiers, show limited effectiveness on MLLMs, while MLLM-specific defenses mainly operate at inference time, filtering suspicious inputs without removing the backdoor embedded in the model. To address this gap and eliminate latent backdoors from MLLMs at their source, we present RACER, a model-level repair framework motivated by a key observation: backdoors induce abnormal layer-to-layer evolution in internal representations, which we term the layer-wise inconsistency anomaly. Importantly, this anomaly is modality-dependent, concentrating primarily in the token region encoding the trigger features that the backdoor model actually relies on. RACER therefore decomposes the fused representation into visual and textual token regions, normalizes their layer-wise inconsistency separately, and recomposes them using modality-aware weights over a deep-layer window, yielding a region-aware inconsistency objective that better captures localized backdoor-induced anomalies. Through a min-max optimization, this objective drives worst-case perturbation synthesis and adversarial fine-tuning against the resulting perturbation to repair the model, suppressing the deep representational directional shifts on which backdoor behaviors rely. RACER requires only 100 clean samples and no knowledge of the trigger, attack objective, or even whether the input model contains a backdoor. Evaluations on three open-source MLLMs across 36 backdoor settings spanning image, text, and multimodal triggers show that RACER reduces the average ASR to 1.1%, reaching 0% in 32 settings, while preserving clean-task utility on both backdoor and clean models.

cs.CR↗
arXiv2026

Rényi Sharpness: A Novel Sharpness that Strongly Correlates with Generalization

Sharpness (of the loss minima) is widely believed to be a good indicator of generalization of neural networks. Unfortunately, the correlation between existing sharpness measures and generalization is not as strong as expected, and sometimes even contradiction occurs. To address this problem, a key observation in this paper is: what really matters for generalization is the average spread (or unevenness) of the spectrum of loss Hessian $\mathbf{H}$. For this reason, conventional sharpness measures, such as trace sharpness $\operatorname{tr}(\mathbf{H})$, which cares about the average value of the spectrum, or max-eigenvalue sharpness $λ_{\max}(\mathbf{H})$, which concerns the maximum spread of the spectrum, are not sufficient to well predict generalization. To characterize the average spread of the Hessian spectrum, we leverage the notion of Rényi entropy in information theory, which captures the unevenness of a probability vector and can thus be extended to a general non-negative vector, such as the Hessian spectrum at loss minima. Specifically, we propose Rényi sharpness, defined as the negative of the Rényi entropy of loss Hessian $\mathbf{H}$. Extensive experiments demonstrate that Rényi sharpness exhibits strong and consistent correlation with generalization in various scenarios. Moreover, two generalization bounds with respect to Rényi sharpness are established by exploiting its desirable reparametrization invariance property. Finally, as an initial attempt to exploit Rényi sharpness for regularization, Rényi Sharpness Aware Minimization (RSAM) is proposed, where a variant of Rényi sharpness is used as the regularizer. RSAM is competitive with state-of-the-art SAM algorithms and far better than conventional SAM based on max-eigenvalue sharpness.

cs.LG↗