中 EN
登录 注册账户

Academic Intelligence · Curated Daily

探索全球前沿学术脉络

AcademicHub 汇聚顶级期刊与预印本平台的实时文献。定制您的专属科研雷达，利用大语言模型自动生成交叉领域文献分析简报。

登录 注册账户
01.
arXiv (CS.LG) 2026-06-17

Verified Detection and Prevention of Concurrency Anomalies in Multi-Agent Large Language Model Systems

作者:
Sajjad Khan↗

arXiv:2606.17182v1 Announce Type: new Abstract: Multi-agent LLM systems share state through memory stores, vector indices, and tool registries. We model such sharing as long-running read-generate-write operations under deterministic-generation semantics – the regime durable-execution engines enforce by deterministic replay – and formalize four concurrency anomalies in TLA+: stale-generation, phantom-tool, causal-cascade, and tool-effect reordering, structural analogues of classical isolation anomalies, each with a TLC counter-example. The exclusion lattice over these anomalies is trivial; the contribution is the mechanically verified realizability and strict separation of one maximal chain within it, $L_0 \subsetneq \cdots \subsetneq L_4$, to our knowledge the first machine-checked consistency hierarchy for such runtimes. A development of 274 Verus obligations (zero assume, zero admit; trust base: two structural axioms and a mutex correspondence) proves the detectors sound and complete against the specifications and each runtime its avoidance set. Three deployed Rust runtimes realize L0-L1 (pessimistic locking, serializable snapshot isolation, default-SI), each verified against stale-generation and refined to its state machine; L2-L4 are exec-mode-verified with dependency-free prevention twins (A3, A6, A2: 0/1000 versus 1000/1000), and L2 is run live across three model families (A3 prevented in all 120 retracted sessions). We reproduce a silent lost update in ByteDance's deer-flow, formalizing its fix as a verified $L_0 \to L_1$ refinement, and exhibit tool-effect reordering in LangGraph's ToolNode on unmodified output, removed by an L3 commit-order sequencer. The verified detector, refinements, and realizability artifacts are the contribution; the phenomena and lattice are classical.

阅读与讨论 → 访问原文 →
02.
arXiv (CS.LG) 2026-06-24

Computational references are not experiments: pre-registered validation of machine-learned sodium-cathode voltages

作者:
Krishna Teja Vepa↗

arXiv:2606.23725v1 Announce Type: cross Abstract: Machine-learning screens for battery materials are trained and judged almost entirely against computed reference voltages, and those references carry their own systematic errors. We report a case in which this matters quantitatively: our own screening stack (a graph-network voltage screen, a prior-art triage layer, and a local PBE+U bench) fails pre-registered validation against experiment-anchored literature values. Verdict thresholds, failure modes, and the primary metric were committed before analysis. On an operator-audited set of known Na-ion cathodes (n = 6 after one documented exclusion; verdict unchanged at n = 7), the raw held-out mean absolute error was 0.67 V, the pre-registered conservative metric, the upper 95% confidence bound of the cross-validated bias-corrected error, was 1.09 V, and the residual was strongly voltage-dependent (r = -0.94), so no additive calibration is valid. On the two compounds where prediction, database reference, and experiment could all be compared, the Materials Project PBE+U reference sat about 0.54 V below measurement: the reference, not the model, dominated the error. A prior-art screen found at least 70% of the targeted Na substitution space already published. We retire the screen, bound what "verified" means for our DFT ledger, and pre-register a calibration audit of it against four benchmark Li couples.

阅读与讨论 → 访问原文 →
03.
arXiv (CS.CL) 2026-06-19

NRITYAM: Language Models Meet Art and Heritage of Dance

作者:
Punit Kumar Singh↗Niladri Ghosh↗Advait Joshi{\i}nst↗Shailee Choudhary↗Michael F\"arber↗Haiqin Yang↗

Language models have become essential tools in shaping modern workflows. However, their global effectiveness hinges on a nuanced understanding of local socio-cultural contexts. To address this gap, we present NRITYAM, a comprehensive benchmark for evaluating the cultural comprehension capabilities of language models in the context of global dance traditions. NRITYAM comprises 9,260 carefully curated question-answer pairs spanning 12 languages, making it the largest dataset dedicated to evaluating cultural knowledge in dance. The dataset has been developed from the ground up through close collaboration with native dance artists and native speakers of the languages, who authored and validated culturally relevant questions specific to their regions. We evaluate a broad set of models, including large language models, small language models, multimodal large language models, and small multimodal language models. As a multilingual and multicultural benchmark, NRITYAM sets a new standard for evaluating the ability of AI systems to understand and reason about traditional performing arts. Detailed dataset samples are available at~\url{https://github.com/niladrighosh03/NRITYAM}.

阅读与讨论 → 访问原文 →
04.
arXiv (CS.LG) 2026-06-16

pFedUL: Layer-Aware Federated Unlearning for Personalized Federated Learning

作者:
Zhuodong Liu↗Xiangyu Li↗Zhihao Zhang↗

arXiv:2606.16304v1 Announce Type: new Abstract: Federated unlearning (FU) enables the removal of specific data contributions from federated learning (FL) models to comply with regulations such as the General Data Protection Regulation (GDPR). However, most existing FU methods are designed for the FedAvg paradigm, where all clients share a single global model. In practice, personalized federated learning (pFL) methods such as FedPer, FedRep, Ditto, and FedBN have become widely adopted due to their superior handling of non-IID data. These methods decompose the model into shared global layers and client-specific personalized layers, fundamentally altering the semantics of unlearning, yet this setting has received little attention. We formalize FU under the pFL paradigm, identifying a tension between unlearning completeness on shared layers and personalization preservation for remaining clients. We then propose pFedUL, a layer-aware selective unlearning framework comprising three components: (1) gradient-based layer-wise contribution attribution that separately quantifies the target client's influence on shared and personalized parameters, (2) adaptive selective unlearning that applies differentiated forgetting strategies across layer types, and (3) a lightweight recalibration protocol enabling remaining clients to restore personalization with minimal overhead. We further introduce two new metrics, Personalization Preservation Score (PPS) and Cross-client Fairness Index (CFI), to evaluate pFL-specific unlearning quality. Experiments on CIFAR-10, CIFAR-100, and FEMNIST under varying non-IID settings indicate that pFedUL achieves unlearning effectiveness comparable to full retraining while maintaining an average of 97.3\% personalized accuracy for remaining clients. Compared with six state-of-the-art FU methods adapted to the pFL setting, pFedUL consistently achieves superior personalization preservation.

阅读与讨论 → 访问原文 →
05.
arXiv (CS.AI) 2026-06-16

LLM Jaggedness Unlocks Scientific Creativity

作者:
Shray Mathur↗J. Anibal Boscoboinik↗Esther H. R. Tsai↗Kevin G. Yager↗

arXiv:2605.10574v3 Announce Type: replace Abstract: As artificial intelligence advances, models are not improving uniformly. Instead, progress unfolds in a jagged fashion, with capabilities growing unevenly across tasks, domains, and model scales. In this work, we examine this dynamic jaggedness through the lens of scientific idea generation. We introduce SciAidanBench, a benchmark of open-ended scientific questions designed to measure the scientific creativity of large language models (LLMs). Given a scientific question, models are asked to generate as many unique and coherent ideas as possible, with the total number of valid responses serving as a proxy for creative potential. Evaluating 19 base models across 8 providers (30 total variants including reasoning versions), we find that jaggedness manifests both across models and within models. First, in a cross-task comparison between general and scientific creativity, improvements in general creativity do not translate uniformly to scientific creativity, revealing divergent capability profiles across models. Second, at the prompt level, stronger models do not improve uniformly; instead, they exhibit high variability, with bursts of creativity on some questions and limited performance on others. Third, at the domain level, individual models display uneven strengths across scientific subfields, reflecting fragmented internal capability profiles. Finally, we show that this jaggedness can be harnessed. We explore mechanisms of inference-time compute, knowledge pooling, and brainstorming to combine models effectively and construct meta-model ensembles that outperform any single model. Our results position jaggedness not as a limitation, but as a resource, a structural feature of AI progress that, when understood and leveraged, can amplify LLM-driven scientific creativity.

阅读与讨论 → 访问原文 →
06.
arXiv (CS.AI) 2026-06-12

The AI Legal Specialist: A Juridically Autonomous Professional Profile for AI Governance

作者:
Nicola Fabiano↗

arXiv:2606.12415v1 Announce Type: cross Abstract: The rapid global expansion of artificial intelligence regulation has generated, across multiple jurisdictions, a demand for legal expertise dedicated to AI that the market has addressed in a fragmented manner. Data protection officers extend their remit beyond data protection law; privacy lawyers reposition themselves toward AI; compliance officers add AI chapters to their existing manuals. This paper argues that none of these adaptive responses adequately covers the professional space opened by the emerging global AI regulatory landscape, of which the EU Artificial Intelligence Act (Regulation (EU) 2024/1689) is the most comprehensive instance, alongside the Council of Europe Framework Convention on AI, the United States executive and sectoral framework, and analogous initiatives in the United Kingdom, Canada, Brazil, China, Japan, Singapore, and beyond. A distinct professional profile is required: the AI Legal Specialist, conceived as a jurist – understood broadly to encompass any professional with advanced legal training – operating at the intersection of legal interpretation and AI governance. The profile is juridically autonomous: it derives its existence from the structure of regulatory obligations generated wherever AI is subject to substantive regulation, rather than from any technical standard or the extension of adjacent roles. The paper provides a juridically grounded definition of the profile, argues for its autonomy from adjacent figures and international standards, proposes a reference competence architecture aligned with the European e-Competence Framework (e-CF, EN 16234-1) as a methodological choice, and articulates the conditions for its operational measurement through key performance indicators. The contribution is intended as a foundation for international standardization of the profile and as a reference for practice, curricula, and adoption across jurisdictions.

阅读与讨论 → 访问原文 →
07.
arXiv (CS.CL) 2026-06-18

CoreMem: Riemannian Retrieval and Fisher-Guided Distillation for Long-Term Memory in Dialogue Agents

作者:
Jiaqi Chen↗Yongqin Zeng↗Shaoshen Chen↗Yijian Zhang↗Hai-Tao Zheng↗Chunxia Ma↗XiuTeng Zhou↗

Personalized dialogue agents require continuous long-term memory to maintain coherent interactions across multiple sessions. However, deploying these capabilities on consumer-grade hardware (e.g., 8 GB VRAM edge devices) introduces severe memory and compute bottlenecks. Existing systems typically rely on isotropic cosine similarity for retrieval and heuristic rules for context compression. These approaches lack a unified theoretical foundation, frequently suffering from the hubness problem in high-dimensional retrieval and syntactic fragmentation during compression. To overcome these limitations, we propose CoreMem, a resource-efficient edge-cloud memory architecture fundamentally unified by information geometry. First, Riemannian retrieval replaces cosine matching with a locally adaptive Fisher-Rao metric, effectively penalizing hub memories via Mahalanobis distance with O(Ndr) Woodbury acceleration for real-time search. Second, Fisher-guided discrete token distillation (FDTD) introduces a hierarchical sentence-to-token compression mechanism. It derives sensitivity scores from Fisher information traces, providing a principled compression-KL tradeoff augmented with explicit structural syntax protection. Evaluated on the LOCOMO and LongMemEval-S benchmarks, CoreMem achieves strong accuracy improvements, yielding substantial gains in Open-domain (+4.51 pp) and Temporal (+4.17 pp) reasoning. Extensive profiling confirms that CoreMem operates seamlessly within a strict 8 GB VRAM budget, successfully bridging the gap between resource-constrained edge devices and the demand for theoretically grounded, lifelong memory agents.

阅读与讨论 → 访问原文 →
08.
Nature (Science) 2026-06-24

Dietary cholesterol activates a Ral-dependent pathway driving LDLR turnover

作者:
Xue Feng↗

Metabolism of the hepatic low-density lipoprotein receptor (LDLR) is a key determinant of cholesterol homeostasis1,2. The molecular switches that coordinate LDLR trafficking and turnover in response to nutritional cues, including high dietary cholesterol, remain poorly defined3–6. Here we identify a new pathway regulated by Ral GTPases that links extracellular cholesterol signals to the intracellular trafficking machinery controlling LDLR turnover. Chronic dietary cholesterol activates the Ral proteins by increasing RAS activity, routing LDLR to lysosomes for degradation and inhibiting its recycling independently of transcriptional regulation or PCSK9. Constitutive activation of Ral via RalGAPB deletion or overexpression of constitutively active Ral mutants in hepatocytes reduces LDLR levels and impairs cholesterol clearance. Ral engages the endocytic RalBP1–REPS1 complex to promote LDLR internalization and lysosomal routing, where LDLR is degraded by the lysosomal protease cathepsin A (CTSA). Ral activation directs CTSA towards lysosomes for maturation while limiting its secretion, further promoting LDLR degradation in lysosomes. Genetic variants in this pathway significantly associate with altered cholesterol in humans. Pharmacological inhibition of CTSA activity increases hepatic LDLR function and improves cholesterol clearance, offering a potential new therapeutic strategy for hypercholesterolaemia and cardiovascular disease. Chronic dietary cholesterol activates Ral GTPases, which promote LDLR internalization and lysosomal degradation through RalBP1–REPS1 and CTSA, thereby reducing cholesterol clearance, whereas CTSA inhibition restores LDLR function and may offer a therapeutic strategy for cardiovascular disease.

阅读与讨论 → 访问原文 →
09.
arXiv (CS.AI) 2026-06-16

FragFuse: Bypassing Access Control of Large Language Model Agents via Memory-Based Query Fragmentation and Fusion

作者:
Zixin Rao↗Wentian Zhu↗Chan Aristella Lu↗Zhaorun Chen↗Wei Niu↗Le Guan↗Bo Li↗Zhen Xiang↗

arXiv:2606.15609v1 Announce Type: cross Abstract: Large language model (LLM) agents increasingly rely on long-term memory to support complex task execution, user personalization, and domain adaptation. Meanwhile, emerging access-control mechanisms for LLM agents are being explored to block policy-violating requests and prevent misuse. We reveal a novel attack surface arising from agent memory operations: prohibited content that would trigger access control can be fragmented across interactions, stored in long-term memory in benign-appearing form, and later reconstructed through memory retrieval without appearing explicitly in the final user query. We propose FragFuse, the first attack that enables unprivileged users to bypass agent access control by exploiting this temporal channel introduced by long-term memory. FragFuse operates in three stages: (1) identifying rejection-responsive fragments via black-box adaptive querying with fragment masking; (2) injecting these fragments into memory using marker carrier queries; and (3) retrieving and fusing the stored fragments through a follow-up attack query. Although FragFuse can be instantiated manually for individual agents, we further develop a surrogate-based optimization scheme that tunes fusion instructions and marker designs, enabling automated attack generation without violating the attacker's threat-model assumptions. We evaluate FragFuse across four representative agent settings and task domains, covering three state-of-the-art agent access-control mechanisms. FragFuse achieves an average bypass success rate of 86.3% and an average end-to-end harmful task success rate of 41.1% across all settings, with only 4.4% average task-success degradation compared with configurations without access control. We also show that alternative defenses, including state-of-the-art prompt-injection detectors and perplexity detectors, do not effectively address this attack.

阅读与讨论 → 访问原文 →
10.
arXiv (CS.LG) 2026-06-17

Continuous-time Optimal Stopping through Deep Reinforcement Learning

作者:
Cosmin Borsa↗Michael Ludkovski↗

arXiv:2606.17545v1 Announce Type: new Abstract: Simulation based solvers for optimal stopping problems must discretize the stopping decision. Under classical dynamic programming, a coarse exercise grid with only a few stopping opportunities can materially undervalue the optimal expected reward, whereas on a very fine grid, approximation errors accumulate through the backward recursion. To remove this limitation, we develop a new reinforcement-learning inspired algorithm that enables us to learn the exercise rule at arbitrarily fine time resolution. Our CARLOS (Continuous-time Adaptive Reinforcement Learning for Optimal Stopping) algorithm utilizes an aggregate deep neural network (ADNN) to learn a joint space-time decision boundary. Starting from a coarse time grid, we progressively increase the frequency of stopping opportunities, while in parallel training the ADNN to refine its timing-value estimates. We moreover design an adaptive sampling strategy that gradually concentrates training effort near the stopping boundary. Benchmarked results show that CARLOS delivers higher prices than existing Bermudan solvers, approaching the American upper bound, and achieves high computational efficiency relative to non-RL comparators.

阅读与讨论 → 访问原文 →
11.
arXiv (quant-ph) 2026-06-24

How rare are Markovian quantum dynamics?

作者:
Charlotte B\"acker↗Nick Maryshchak↗Walter T. Strunz↗

arXiv:2606.24511v1 Announce Type: new Abstract: A profound understanding of decoherence and dissipation in quantum dynamics is crucial for the realistic modeling of the evolution of quantum systems. In open quantum dynamics one distinguishes between a memoryless, so-called Markovian evolution and dynamics incorporating memory effects, termed non-Markovian. In this work we study how prevalent memory effects are in the set of all such dynamics. We thus investigate how often a Markovian description is applicable. This question is approached by investigating randomly generated two-step qubit dynamics with respect to different concepts and witnesses of non-Markovianity. We observe that almost all dynamics are non-Markovian, and only a small (yet finite) fraction is Markovian. Furthermore, we study how this proportion changes when considering certain subclasses such as lower rank or mixed-unitary dynamics. Importantly, our results shed light on the relative ratios of – and interrelations between – the sets of dynamics that are non-Markovian with respect to different criteria. Finally, we investigate the fraction of dynamics in which the memory effects are necessarily of quantum nature and establish a connection between two recently developed concepts that characterize the quantumness of memory in non-Markovian dynamics.

阅读与讨论 → 访问原文 →
12.
arXiv (CS.CV) 2026-06-15

FBSDiff++: Improved Frequency Band Substitution of Diffusion Features for Efficient and Highly Controllable Text-Driven Image-to-Image Translation

作者:
Xiang Gao↗Yunpeng Jia↗

With large-scale text-to-image (T2I) diffusion models achieving significant advancements in open-domain image creation, increasing attention has been focused on their natural extension to the realm of text-driven image-to-image (I2I) translation, where a source image acts as visual guidance to the generated image in addition to the textual guidance provided by the text prompt. We propose FBSDiff, a novel framework adapting off-the-shelf T2I diffusion model into the I2I paradigm from a fresh frequency-domain perspective. Through dynamic frequency band substitution of diffusion features, FBSDiff realizes versatile and highly controllable text-driven I2I in a plug-and-play manner (without need for model training, fine-tuning, or online optimization), allowing appearance-guided, layout-guided, and contour-guided I2I translation by progressively substituting low-frequency band, mid-frequency band, and high-frequency band of latent diffusion features, respectively. In addition, FBSDiff flexibly enables continuous control over I2I correlation intensity simply by tuning the bandwidth of the substituted frequency band. To further promote image translation efficiency, flexibility, and functionality, we propose FBSDiff++ which improves upon FBSDiff mainly in three aspects: (1) accelerate inference speed by a large margin (8.9$\times$ speedup in inference) with refined model architecture; (2) improve the Frequency Band Substitution module to allow for input source images of arbitrary resolution and aspect ratio; (3) extend model functionality to enable localized image manipulation and style-specific content creation with only subtle adjustments to the core method. Extensive qualitative and quantitative experiments verify superiority of FBSDiff++ in I2I translation visual quality, efficiency, versatility, and controllability compared to related advanced approaches.

阅读与讨论 → 访问原文 →
13.
arXiv (CS.CL) 2026-06-16

SkillWiki: A Living Knowledge Infrastructure for Agent Skills

作者:
Dingcheng Huang↗Yuda Ding↗Bingshuo Liu↗Qingbin Liu↗Xi Chen↗Jiang Bian↗Hongliang Sun↗Zhiying Tu↗Dianhui Chu↗Xiaoyan Yu↗Dianbo Sui↗

While knowledge is managed through Wikipedia and software through GitHub, agent skills still lack an infrastructure for large-scale production, governance, and evolution. SkillWiki is a living knowledge infrastructure that supports the organization, grounding, and continuous evolution of agent skills by transforming heterogeneous knowledge into reusable skill assets linked to their originating evidence. Our demonstration presents the complete skill lifecycle, from knowledge ingestion and skill production to provenance-aware exploration, governance, and execution-driven evolution. SkillWiki highlights a future in which knowledge, skills, and execution experience co-evolve within a shared infrastructure. The live demonstration and source code are publicly available at https://github.com/Huangdingcheng/SkillWiki.

阅读与讨论 → 访问原文 →
14.
arXiv (quant-ph) 2026-06-17

Variational Quantum Eigensolver-Based Quantum Bootstrap Embedding for Molecules

作者:
Derek Peng↗

arXiv:2606.17095v1 Announce Type: cross Abstract: Simulating strongly correlated molecular systems on near-term quantum hardware remains challenging due to modern hardware's limited quantum volume and moderate-fidelity qubits. One potential way to circumvent this challenge is through bootstrap embedding (BE). Bootstrap embedding breaks molecules into smaller fragments that are then embedded into the "bath" of other fragments in an iterative way. Bootstrap embedding is appealing for quantum simulation because fragmenting the system reduces the qubit requirements for any given fragment. In this work, we develop a quantum bootstrap embedding (QBE) workflow that uses variational quantum eigensolver (VQE) fragment solvers and study the algorithmic choices that determine the overall VQE-QBE algorithm's success. To improve efficiency, we introduce FastAdaptVQE, a sparse matrix-accelerated form of the adaptive variational quantum eigensolver (ADAPT-VQE) that replaces symbolic commutator evaluation with direct statevector linear algebra, and MatrixFreeAdaptVQE, a matrix-free extension that removes the sparse-matrix memory bottleneck that appears when treating larger fragments. We also modify the ADAPT-VQE operator selection step by replacing the purely greedy choice with a look-ahead strategy. Benchmarks on $H_4$ and $F_2$ reach chemical accuracy, within 1 kcal/mol of bootstrap embedding results using a full configuration interaction (FCI) solver. These results show that combining QBE with VQE can accurately calculate energies of molecular systems. This research lays the foundation for extending energy calculations to larger molecular systems and quantum materials on near-term quantum hardware.

阅读与讨论 → 访问原文 →
15.
arXiv (CS.CV) 2026-06-19

Single-Stage Hierarchical Rectification for Weakly Supervised Histopathology Segmentation

作者:
Duc T. Nguyen↗Hoang-Long Nguyen↗Thanh-Ha DO↗Huy-Hieu Pham↗

Existing weakly supervised semantic segmentation (WSSS) methods in computational pathology rely on a multi-stage paradigm: class activation map (CAM) generation, offline pseudo-mask refinement, and fully supervised retraining. While established, this decoupled approach presents fundamental limitations. The multi-stage process not only incurs high computational training costs but also suffers from error propagation: local texture biases in shallow CNN layers generate false-positive artifacts that subsequent refinement steps often fail to correct. To address these persistent challenges through a simple yet highly effective approach, we propose the Single-Stage Hierarchical Rectification (SSHR) framework. Rather than passively refining CAMs post-hoc, our method proactively purifies intermediate feature representations during the forward pass. We introduce a Hierarchical Feature Rectification Module (HFRM) that utilizes deep global semantic context to filter out local anomalies in shallow layers. This mechanism generates high-fidelity activation maps directly within a single training loop. Experiments on the LUAD-HistoSeg and BCSS datasets demonstrate that SSHR outperforms state-of-the-art multi-stage methods. Furthermore, SSHR reduces training duration by 2 to 5 times. This efficiency minimizes computational overhead and accelerates clinical translation for large-scale histopathology workflows. The code is available at: https://github.com/trongduc-nguyen/SSHR

阅读与讨论 → 访问原文 →
16.
arXiv (quant-ph) 2026-06-12

The table maker's quantum search

作者:
Benjamin C. A. Morrison↗Stefanos Kourtis↗

arXiv:2601.13306v2 Announce Type: replace Abstract: We show that quantum search can be used to compute the hardness to round an elementary function, that is, to determine the minimum working precision required to compute the values of an elementary function correctly rounded to a target precision of $n$ digits for all possible precision-$n$ floating-point inputs in a given interval. For elementary functions $f$ related to the exponential function, quantum search takes time $\tilde O(2^{n/2} \log (1/\delta))$ to return, with probability $1-\delta$, the hardness to round $f$ over all $n$-bit floating-point inputs in a given binade. For periodic elementary functions in large binades, standalone quantum search yields an asymptotic speedup over the best known classical algorithms and heuristics. We then estimate the resources required for a fault-tolerant implementation of the proposed algorithm for the $\sin$ and $\cos$ functions in double precision. We find that, although the algorithm can in principle compete with the fastest known practical method for computing the hardness to round over all binades in the format, it requires qubit coherence times that are unrealistically long for present technology.

阅读与讨论 → 访问原文 →
18.
arXiv (CS.AI) 2026-06-17

How Inference Compute Shapes Frontier LLM Evaluation

作者:
Jessica McFadyen↗Ole Jorgensen↗Harry Coppock↗Kevin Wei↗Cozmin Ududec↗

arXiv:2606.17930v1 Announce Type: new Abstract: AI evaluations are shifting toward harder tasks that benefit from longer trajectories involving tool use and iterative problem solving. As a result, performance is increasingly sensitive to the amount and allocation of compute available at test time ("inference compute"). Yet many evaluations still report performance at a single restrictive budget, meaning that low scores may reflect the evaluation setup rather than the model's underlying capability. To test this, we evaluate up to 12 frontier language models on seven challenging benchmarks spanning software engineering, mathematics, medicine, and cybersecurity. We use a controlled setup combining three simple inference-scaling interventions: larger token budgets, context compaction, and repeated submission attempts, guided either by the model itself or by minimal correctness feedback. We find three main results. First, larger token budgets substantially improve performance on benchmarks across multiple domains, including cybersecurity, FrontierMath, Humanity's Last Exam, and TerminalBench. Second, fixed-budget evaluations can increasingly understate frontier capability as models advance. Newer models reach higher performance at large budgets, where they unlock harder tasks and solve them more reliably. Third, benchmarks differ in which inference-scaling methods help most: repeated submission broadly improves performance, but the value of larger token budgets, external feedback, and parallel attempts varies by benchmark. Overall, our results show that benchmark scores are protocol-dependent. We therefore argue that evaluations should report capability as a function of inference-time compute, specify protocol choices explicitly, and compare model generations over a large shared compute range at matched budgets, especially in safety- or policy-relevant settings.

阅读与讨论 → 访问原文 →
19.
arXiv (CS.LG) 2026-06-16

Size Doesn't Matter: Cosine-Scored Sparse Autoencoders

作者:
Silen Naihin↗Lev Stambler↗

arXiv:2606.15054v1 Announce Type: new Abstract: Sparse autoencoders (SAEs) detect features via inner product, so a feature's activation scales with both its directional alignment and the input's norm. Under BatchTopK, high-norm tokens inflate all pre-activations simultaneously, claiming dictionary slots regardless of content alignment. This matters because sublayer normalization has already discarded the magnitude the score measures, so the encoder detects a quantity the model does not read. We replace the score with a learned blend of cosine similarity and input magnitude, letting the optimizer choose how much norm to use; a per-feature extension lets each feature decide independently. In both regimes, training is free to recover inner product but never does, with no feature ever choosing more than half-magnitude dependence. At matched reconstruction, the cosine encoder learns features that align with human-recognizable concepts far more often than standard, filling dictionary slots that inner product wastes on norm detectors. Loss reweighting that equalizes gradients barely closes the gap, confirming forward-pass score geometry as the lever. The advantage is not universal across tasks or depths, but we believe cosine scoring should be the default for dictionary learning on normalized representations.

阅读与讨论 → 访问原文 →
20.
arXiv (CS.AI) 2026-06-11

Erased but Not Forgotten: How Backdoors Compromise Concept Erasure

作者:
Tobias Braun↗Jonas Henry Grebe↗Marcus Rohrbach↗Anna Rohrbach↗

arXiv:2504.21072v3 Announce Type: replace-cross Abstract: The expansion of text-to-image diffusion models has raised concerns about harmful outputs, from fabricated depictions of public figures to sexually explicit imagery. To mitigate such risks, prior work has proposed concept erasure methods that aim to sever unwanted concepts from the model via fine-tuning, yet it remains unclear whether these approaches truly remove all links to the harmful concept or merely conceal superficial connections. In this work, we reveal a critical vulnerability, the Erasure Evasion Backdoor (EEB): an adversary binds a backdoor trigger to a concept slated for removal, and this malicious link survives subsequent erasure. We show that both black-box and white-box adversaries can instantiate this threat. Across six state-of-the-art erasure methods, including robust ones that explicitly search for alternative representations of the target concept, EEB consistently exposes harmful content: up to 82% success against celebrity-identity unlearning, up to 94% for object erasure, and up to 16 times amplification of explicit-content exposure. While EEB uncovers a blind spot in current erasure methods, it also provides a diagnostic tool for stress-testing future concept erasure techniques.

阅读与讨论 → 访问原文 →
21.
arXiv (CS.CL) 2026-06-11

Redesign Mixture-of-Experts Routers with Manifold Power Iteration

作者:
Songhao Wu↗Ang Lv↗Ruobing Xie↗Yankai Lin↗

Router is the cornerstone component to the Mixture-of-Experts models. Serving as expert proxies, the rows of the router matrix compute their similarity to the MoE inputs to determine which subset of experts is activated. Ideally, each router row is designed to encode the expert matrix into this representative vector, such that its dot-product with token can better reflect token-expert affinity. However, there exists no design principles to enforce this condensation. In this paper, we propose to align each router row with the principal singular direction of the associated expert, as this direction provides the most expressive mathematical description of a matrix. Based on this principle, we propose a router redesign with Manifold Power Iteration (MPI). Specifically, it introduces a "Power-then-Retract" paradigm, where a power iteration step is performed on the router weights, followed by a retraction to impose a norm constraint to ensure both efficiency and stability. Theoretically, we show that MPI drives router rows to converge toward the principal singular directions of associated experts. Empirically, we pretrain MoE model across scales from 1B to 11B parameters to confirm that this alignment facilitates more effective MoE models.

阅读与讨论 → 访问原文 →
22.
arXiv (CS.CL) 2026-06-17

FeedEval: Pedagogically Aligned Evaluation of LLM-Generated Essay Feedback

作者:
Seongyeub Chu↗Jongwoo Kim↗Munyong Yi↗

Going beyond the prediction of numerical scores, recent research in automated essay scoring has increasingly emphasized the generation of high-quality feedback that provides justification and actionable guidance. To mitigate the high cost of expert annotation, prior work has commonly relied on LLM-generated feedback to train essay assessment models. However, such feedback is often incorporated without explicit quality validation, resulting in the propagation of noise in downstream applications. To address this limitation, we propose FeedEval, an LLM-based framework for evaluating LLM-generated essay feedback along three pedagogically grounded dimensions: specificity, helpfulness, and validity. FeedEval employs dimension-specialized LLM evaluators trained on datasets curated in this study to assess multiple feedback candidates and select high-quality feedback for downstream use. Experiments on the ASAP++ benchmark show that FeedEval closely aligns with human expert judgments and that essay scoring models trained with FeedEval-filtered high-quality feedback achieve superior scoring performance. Furthermore, revision experiments using small LLMs show that the high-quality feedback identified by FeedEval leads to more effective essay revisions. We release our code and curated datasets at: https://github.com/BBeeChu/FeedEval.git.

阅读与讨论 → 访问原文 →
23.
arXiv (CS.AI) 2026-06-24

The Measurable Majority

作者:
Lawrence S. Moss↗Arthur Paul Pedersen↗

arXiv:2606.23853v1 Announce Type: cross Abstract: This paper studies strict majority reasoning in finite electorates using so-called $social decision frames$: finite sets of voters equipped with distinguished families of coalitions interpreted as those voting blocs evaluated to form a strict majority. A coherence criterion for qualitative majority judgments is identified and shown to give an exact characterization for representability of strict majorities by finitely additive measures. In addition, a minimal natural logic for reasoning about strict majorities is shown to be sound and complete. These developments motivate examination of associated combinatorial questions concerning incoherence in finite families of sets; partial results and a conjecture are given. Finally, the results of this paper are applied to correct a classical representation theorem for weak qualitative probability structures due to Patrick Suppes and to establish a May-type characterization for ordinary strict majority rule for social decision frames.

阅读与讨论 → 访问原文 →
24.
arXiv (CS.LG) 2026-06-16

Tangram: Unlocking Non-Uniform KV Cache Compression for Efficient Multi-turn LLM Serving

作者:
Hyungmin Kim↗Minsoo Kim↗Hongseok Kim↗Jungwook Choi↗

arXiv:2606.06302v2 Announce Type: replace Abstract: Multi-turn LLM serving accumulates dialogue history whose Key-Value (KV) cache grows with every turn and every user, quickly exceeding the model weights themselves and making memory – not compute – the binding constraint on throughput. Non-uniform KV compression, which allocates heterogeneous budgets across attention heads, preserves accuracy far better than uniform schemes, yet remains impractical: modern serving stacks assume identical KV lengths across heads, so heterogeneity traps freed memory as page fragmentation, spends up to 25% of prefill time reclaiming scattered pages, and skews GPU workloads that inflate decode latency by up to $1.7\times$ or burn 15–20% of each decode step on re-planning. We observe that this heterogeneity need not be discovered at runtime: head-wise retention follows a two-level structural regularity – an input-invariant head ranking with narrowly bounded per-head ratios – that can be calibrated offline from as few as 50 samples. Building on this insight, we present Tangram, a serving framework that statically resolves what prior systems handle dynamically: Budget Reservation fixes each head's post-compression footprint at scheduling time, eliminating page reclamation; Ragged Paging clusters similar-budget heads into independent page tables, turning fragmentation into reclaimable memory; and Ahead-of-Time Load Balancing precomputes balanced GPU partitions with zero runtime planning. Implemented on vLLM, Tangram serves as a drop-in substrate for existing non-uniform compression methods, matching their accuracy while improving end-to-end throughput by up to $2.6\times$ over the full-KV baseline. Our implementation is publicly available at https://github.com/aiha-lab/TANGRAM.

阅读与讨论 → 访问原文 →
25.
medRxiv (Medicine) 2026-06-22

Characteristics and Outcomes of Gene-Elusive Dilated Cardiomyopathy

作者:
Cannie↗Bakalakos↗Syrris↗Protonotarios↗Lorenzini↗Guttmann↗O'Mahoney↗Savvatis↗Sekhri↗Mohiddin↗S. A↗Kaski↗…

Background and Aims Genetic testing in dilated cardiomyopathy (DCM) guides risk stratification and family screening. Likely pathogenic or pathogenic (LP/P) variants are identified in approximately one-third of patients, leaving many without a genetic diagnosis. Cohort studies suggest that "gene-elusive" patients have a lower risk of adverse events. This study aims to better characterise this group and identify factors associated with adverse outcomes. Methods Consecutive and unrelated DCM patients undergoing genetic testing and returning no LP/P variants were retrospectively recruited and compared to two control cohorts of DCM patients carrying LP/P variants in LMNA and TTN for a primary composite endpoint of end-stage heart failure (ESHF) or malignant ventricular arrhythmia (MVA). Results Among patients without prior MVA, the composite endpoint occurred in 36/423 (8.5%) gene-elusive, 14/39 (35.9%) LMNA and 11/100 (11%) TTN cardiomyopathy patients (log-rank p

阅读与讨论 → 访问原文 →