Shai Nahum Gefen, Vered Chassida, Dmitry Basin, Guy Feigenblat
原文摘要
Government intelligence teams need LLM agents that can analyze large, messy evidence collections, but security and cost constraints often rule out hosted frontier models and broad, hard-to-audit tool surfaces. We report CLEO: Constrained Loop for Environment Optimization, a method for fitting a minimal agent harness to a fixed, self-hostable LLM. Given an expert-validated dataset, CLEO admits only the tools, guardrails, memory policies, and loop components that improve accuracy while remaining general and auditable. On our held-out dataset, the CLEO harness reaches an 83.8% pass rate, within 1.2 points of the strongest out-of-the-box Claude Code SDK baseline, while being two orders of magnitude smaller.
Gino Kwun, Dhanvi Bharadwaj, Gokul Subramanian Ravi
原文摘要
Variational Quantum Algorithms (VQAs) potentially offer a pathway to practical quantum advantage, but their optimization is heavily hindered by barren plateaus and numerous local minima. While classically simulable Clifford circuits can warm-start VQAs to accelerate convergence, existing heuristic-based initialization methods struggle to scale within vast combinatorial search spaces. We address this bottleneck with CRiSP (a Clifford Reinforcement Learning agent for State Preparation), which casts Clifford prefix selection as a sequential decision-making problem. CRiSP utilizes Neural-Guided Monte Carlo Tree Search, driven by a Transformer-based policy trained via self-play, to insert learned Clifford gates before fixed parameterized rotations. This enables the construction of high-quality initial states entirely through polynomial-time classical stabilizer simulation. By integrating a curriculum learning strategy that progressively expands the search horizon, the agent efficiently scales to deep circuits. Evaluated on QAOA benchmarks of up to $22$ qubits and $1{,}370$ parameters, CRiSP outperforms state-of-the-art Clifford initialization methods by a mean of $3.19\times$ (max $45.02\times$) in average energy accuracy and $2.41\times$ (max $16.01\times$) in best-achieved energy accuracy. Assessments on VQE tasks further demonstrate the framework's robustness and generalizability.
Sotiris Chatzimiltis, Mahdi Boloursaz Mashhadi, Mohammad Shojafar, Rahim Tafazolli
原文摘要
Next-generation open radio access network (Open RAN) and AI-RAN deployments turn the RAN into a disaggregated, compute-aware platform whose service continuity depends on many control loops acting at different timescales. Resilience in this setting is best seen as a lifecycle spanning anticipation, absorption, adaptation, recovery, and evolution. Yet, existing mechanisms remain stage-specific and are rarely coordinated across the control hierarchy. This paper argues that agentic AI offers an orchestration layer that coordinates this lifecycle, turning operator intent into coordinated action while keeping slow language-model reasoning off the fast control path. Using mass-attachment signaling storms as a running study, we build a discrete-event Open RAN attach simulator that captures retry amplification, shared-compute contention, and non-instantaneous provisioning, and we propose a three-tier framework in which an orchestrator at the service management and orchestration layer sets a standing policy, a per-site non-real-time agent judges storms and anticipates scheduled crowds, and a deterministic near-real-time loop enforces every decision. Across experiments, the agentic controller approaches the resilience of a fully provisioned pool of servers at a quarter of its capacity by reading a scheduled event from a calendar and provisioning ahead of it. No reactive weight setting recovers this once the surge arrives. Free-text operator intents ground into control directives at 94% exact match on previously unseen intents, and a single instruction reshapes live behavior end to end. We also quantify the latency, cost, and contention limits that bound it. The supplementary material provides analytical proof of the derived Theorem on anticipation horizon, as well as the subsequent Corollary results on reactive provisioning limits.
Technology-mediated audience participation (TMAP) enables audiences to influence live performances, yet balancing interaction with meaningful agency remains challenging. While AI-generated music and MR enable new forms of audio and visual audience participation, little is known about how these participation modalities influence audience experiences during live musical performance. In particular, it remains unclear how they shape perceived agency, collaboration, and engagement, and how they complement one another within TMAP systems. We present CrowdForge, an AI- and MR-mediated paradigm where audiences collectively control generative music and virtual environments during a live performance featuring a guitarist and DJ. Through a mixedmethods concert study evaluating three interaction conditions (audio, visual, and combined audio-visual), we demonstrate that observability is the primary driver of audience agency. While visual interaction provided easily observable feedback, audio parameters offered macro-structural control over the music. Audiences tended to use the control enabled by AI audio generation to test musical boundaries, introducing a productive adversarial tension between the crowd and the musicians. This dynamic pushed creative limits, deepened engagement, and generated moments of musical coherence when performers successfully resolved audience-induced challenges. We argue future TMAP systems should prioritize human-centered, highly observable interactions to build agency, while combining the expressiveness of music with the clarity of visual feedback.
Medical imaging has served as primary proving ground for clinical AI, yet a decade of intense research has not translated into proportionate bedside impact. We argue that this gap is not primarily a product of insufficient algorithmic performance, inadequate regulation, or limited explainability. Rather, it reflects a structural misalignment, between how AI systems are designed and evaluated, and how clinical decisions are made. This work identifies 8 interconnected dimensions of this misalignment: 1) the dominance of pixel-only models in a multimodal clinical world; 2) the erosion of physician trust through opaque and inflexible systems; 3) the unfulfilled promise of foundation models in data-sparse medical domains; 4) the persistent bottleneck of non-shareable under-curated datasets; 5) the gap between validated algorithms and deployable clinical platforms; 6) the slow, unmonitored translation of evidence into routine practice; 7) the risk of AI displacing rather than extending physician judgment; and 8) the failure of prediction-centric AI to generate actionable, agentic clinical guidance. For each dimension, we reframe the problem and propose a path forward, culminating in a vision of physician-centered AI solutions with agentic reasoning that extend, rather than replace, clinical judgment.
Mohammad Kachuee, Phu Mon Htut, Sheng Li, Nikolaos Pappas, Miguel Ballesteros, Yassine Benajiba, Sujith Ravi, Dan Roth
原文摘要
Agentic AI extends large language models (LLMs) from generating responses to planning and executing complex workflows, substantially expanding their utility while exposing applications to unsafe actions, information leakage, and adversarial manipulation. Multi-agent systems further expand these capabilities through specialization and collaboration, but distributed authority, heterogeneous roles, and opaque or potentially compromised participants make safeguarding even more challenging. Existing defenses address important individual threats but provide fragmented coverage across local execution, system-level interactions, and enforcement mechanisms. We introduce MAS-Guard, a unified framework that closes this gap for safeguarding single- and multi-agent AI applications containing both instrumentable and opaque agents. Within instrumentable agents, the proposed local Agent Guardrail (A-Guard) layer applies task-specific experts to assess agentic events. At the system level, developer-owned Agent Charters and a Control Plane combine deterministic structural checks of routes, agent capabilities, and message formats with semantic assessments of roles and message content. These complementary layers enable enforcement through a shared set of interventions: allow, reflection, confirmation, and block. Evaluated across six benchmarks covering diverse single- and multi-agent scenarios with three agent LLMs from different providers, MAS-Guard reduces the overall attack success rate by 71.5-86.3% relative to unguarded agents, while meeting the utility reference criterion on benign tasks.
Best-of-N decoding improves generation quality by sampling multiple candidates and selecting the best, but it allocates equal compute to sources regardless of how much they benefit from additional sampling: the expected benefit is small for sources with low-variance score distributions and potentially much larger for those with high-variance distributions. To address this shortcoming, we propose LAND (Learning an Adaptive Number of Draws), which dynamically allocates samples according to their posterior-predictive marginal value. LAND learns a discrete empirical mixture of source-conditioned score distributions from calibration data, updates its belief about a source as completion scores are observed, and stops sampling when the expected improvement of another draw falls below a shared compute price. Across machine-translation experiments with multiple models and language pairs, LAND achieves better MetricX scores than fixed best-of-N at matched average compute. Conversely, at matched translation quality, LAND requires approximately 20-40% fewer candidate generations. Adaptive allocation therefore provides a simple way to retain the benefits of best-of-N while substantially improving its compute efficiency.
Eric Benhamou, Jean-Jacques Ohana, Beatrice Guez, Ethan Setrouk, Chamyl Saadi, Thomas Jacquot
原文摘要
Financial research agents can generate and monitor trading rules faster than a human can document the resulting selection process. Candidates appear at data-dependent times, weak rules are abandoned, and promising ones are watched until they cross a reporting threshold. We introduce PROMOTE, a protocol that assigns every candidate predictable testing wealth at birth and accumulates a cost-aware e-process only from subsequent returns. If lifetime allocations sum to at most one, a weighted Ville argument bounds the probability of ever promoting a rule whose declared bounded net score has always had nonpositive conditional mean. The result permits dependent candidates, adaptive births, and indefinite monitoring. We audit the implementation in two settings. A retrospective pseudo-out-of-sample replay contains 24 panel--rule hypotheses, 3,000 shared-sign null paths, injected alternatives, and public ETF and industry data. Repeated unadjusted tests produce a family crossing in 94.5% of null paths, versus 0.0% for PROMOTE; at a 15-bp aligned ETF-component perturbation, PROMOTE has 70.1% target power and median crossing session 2,327. A separate exact mean-zero experiment lets an agent select candidate identities and birth times from past scores; the funded gate has 1.77% false-family crossings across 3,000 programs. The market replay produces zero promotions after costs. The certificate controls historical false promotion under the declared null; it is not a certificate that an earlier edge remains profitable at deployment. AI-use disclosure. Generative AI tools, including OpenAI Codex, assisted manuscript preparation. For this preprint version, Codex assisted with LaTeX editing, author and venue metadata, compilation, and layout checks.
As multimodal conversational systems increasingly engage in spoken interaction, their ability to navigate paralinguistic social cues has become a critical bottleneck for natural human-AI communication. However, existing evaluations of machine emotional intelligence assess reasoning exclusively through isolated text or passive acoustic perception, overlooking the complex cross-modal reasoning required for active, multi-turn dialogue. We introduce \textsc{SpeechEQ}, a comprehensive framework designed to evaluate the sociolinguistic reasoning of Speech-Language Models (SLMs). The framework includes a validated dataset of 2,265 dialogues across 15 Emotional Quotient (EQ) subscales grounded in EQ-i 2.0 theory, along with a multi-turn evaluation protocol measured by our proposed Spoken EQ (SEQ) score inspired by human EQ assessments. Experiments show limitations in how both existing Speech Emotion Recognition and end-to-end Speech-Language Models understand and apply paralinguistic cues through speech. While end-to-end architectures outperform cascaded systems, \textsc{SpeechEQ} reveals that current multimodal models remain bottlenecked by a text-reliant ``modality shortcut,'' an alignment-induced ``safety trap,'' and ``contextual amnesia,'' highlighting the barriers to truly emotionally aware AI. Our benchmark can be accessed at https://huggingface.co/datasets/SpeechEQ/SpeechEQ and demo page at https://speecheq.github.io
Daehwan Ahn, Abdullah Almaatouq, Monisha Gulabani, Kartik Hosanagar
原文摘要
This paper bridges the gap in Human-Computer Interaction (HCI) research by comparatively assessing the effects of interpretability and outcome feedback on user trust and collaborative performance with AI. Through novel pre-registered experiments (N=1,511 total participants) using an interactive prediction task, we analyzed how interpretability and outcome feedback influence users' task performance and trust in AI. The results counter the widespread belief that interpretability drives trust, showing that interpretability led to no robust improvements in trust and that outcome feedback had a significantly greater and more reliable effect. However, both factors had modest effects on participants' task performance. These findings suggest that (1) interpretability may be less effective at increasing trust than factors like outcome feedback, and (2) augmenting human performance via AI systems may not be a simple matter of increasing trust in AI, as increased trust is not always associated with equally sizable performance improvements. Our exploratory analyses further delve into the mechanisms underlying this trust-performance paradox. These findings present an opportunity for research to focus not only on methods for generating interpretations but also on techniques that ensure interpretations impact trust and performance in practice.
Every multimodal-LLM multiple-choice benchmark ends with a scoring rule that converts a model's response, often free-form text rather than a clean option letter, into a right/wrong verdict, buried in its code. Unlike the dataset, this rule is chosen without justification and left unexamined, yet it sets the published number. We reproduce the official scorers of eight peer-reviewed benchmarks on five open vision-language models (VLMs) that fit a single 16 GB GPU, and introduce a blinded human-plus-LLM-judge adjudication (Cohen's kappa=0.956) that recovers what the model actually answered. The scoring rule alone moves accuracy by up to 22 points on identical outputs, yet exactly 0 wherever a model emits clean option letters, so it hides on strong-model leaderboards. It is the fallback and not the rambling that does the damage: the benchmark with the highest non-compliance we measure has an artifact of exactly zero, because its scorer abstains instead of guessing; a reasoning model with non-compliance 0.000 still swings 9.5 points with which regex match of its trace the parser keeps. No standard rule is reliable: snapping a non-answer to its closest option credits answers never given yet buries correct ones, abstention discards answers stated in words, and likelihood scoring follows answer-token priors, so a single grade conflates capability with luck rather than measuring either. We release all code, labels, and the protocol, and argue benchmarks should report a non-compliance rate and at least two rules, not one unexamined grade.
Olalekan David-Banjo, Chinazunwa Uwaoma, Paul Witman
原文摘要
Crime-reporting mobile applications can enhance community safety and engagement in urban areas, yet their adoption remains low, partly because developers prioritize technological features over residents' social realities. This study used an exploratory sequential mixed-methods design to identify and validate the sociotechnical factors that drive adoption. Two focus groups and an analysis of 660 app reviews revealed that users' initial sense of usefulness fades, whereas sustained engagement depends on perceived value (PV), comprising six dimensions: safety, anonymity, recency, autonomy, camaraderie, and accessibility. A survey of 505 urban residents, analyzed with PLS-SEM, validated PV as a formative higher-order construct formed by four of these dimensions (safety, anonymity, autonomy, and camaraderie). PV was the strongest predictor of behavioral intention, ahead of effort expectancy, facilitating conditions, and performance expectancy, and behavioral intention strongly predicted actual use. The model explained 75% of the variance in behavioral intention and 51% in actual use. The findings offer design principles for community-centered crime-reporting technologies.
Recent language model (LM) safety failures have largely been studied as disparate phenomena. In agentic misalignment, a model blackmails an executive to avoid being replaced. In alignment faking, it complies with harmful requests during training to keep its values from being modified. We argue that these behaviours share a common decision structure: a model assigned a goal must either take a harmful action to preserve it or refuse and lose it. We formalize this as a goal-harm conflict and present the same conflict under various elicitation protocols that range from compact forced-choice prompts to simulated agentic environments. These protocols span two perspectives on LM decision-making, which we term (i) the logit view, reading preferences from answer-token probabilities, and (ii) the sampling view, classifying open-ended responses with an LM judge. Under the logit view, we show that goal importance and harm severity additively shape how a model resolves these value conflicts. Under the sampling view, we derive two factorizations of the harmful-action probability that safety evaluations commonly report. The first measures how well a preference read from answer tokens carries over into behaviour, such as making a harmful tool call. The second attributes a low harmful-action rate either to refusal or to responses that reach no resolution. We then show how fictional context changes a model's preference under the logit view, and how tool use and partial observability -- two features of agentic evaluations -- change its harmful-action rate under the sampling view. Our work offers a descriptive framework for studying how LMs resolve such value conflicts and for estimating low-probability harmful actions at a finer granularity than the aggregate harm rates commonly reported in safety evaluations.
Sewoong Lee, Marc E. Canby, Ikhyun Cho, Julia Hockenmaier
原文摘要
The term "linear representation hypothesis" (LRH) has appeared across diverse subfields of artificial intelligence, neuroscience, and cognitive science. But previous works have not consistently treated the LRH as a falsifiable scientific hypothesis; we analyze these inconsistencies and examine their implications for how prior theoretical and methodological results should be interpreted. Based on this analysis, we argue that claims regarding linear representations become well-defined only through careful examination of the model, representation location, feature definition, and evaluation dataset. We therefore propose a more rigorous formalization of the LRH that makes these dependencies explicit and allows the hypothesis to be evaluated as a falsifiable scientific claim. Finally, we identify some non-trivial open problems that warrant further attention from the research community.
Sheridan Feucht, Benno Krojer, Sarah Wang, Henry Abrahamsen, Byron C. Wallace, David Bau
原文摘要
How do VLMs map from pixels to semantics? To understand this general question, we focus on a narrow one: studying how VLMs perform optical character recognition (OCR). Across four models, we identify attention heads causally necessary for OCR, and discover that these are in fact general-purpose heads that output interpretable semantic features across all image tokens. For example, pointing these heads at an image token containing the word "bike" causes Qwen3-VL-8B to output "bike," but pointing them at a bird wing causes the model to output the token "feathers." We collapse these heads' attention weights into a single verbalization lens transformation that reveals interpretable semantic features in hidden states across all layers. When combined with projection to vocabulary space, we can obtain interpretable labels starting from layer 0, showing that image representations are in fact aligned with language in early layers. We find that we can also use the inverse of this transformation to edit non-word concepts, e.g., replacing a tractor with a revolver in a naturalistic image, providing causal evidence that this subspace is useful for more than just OCR. Our results are an example of how the study of specific mechanisms can shed light on broader interpretability problems.
Large Language Model (LLM) agents excel at automating data science workflows, yet often underutilize rigorous statistical methods in R. Existing Retrieval-Augmented Generation (RAG) approaches rely on semantic similarity, ignoring the data-distribution and structural characteristics crucial for determining a statistical method's applicability. To address this, we propose DARE (Distribution-Aware Retrieval Embedding), a representation learning framework incorporating these data characteristics into R function retrieval. Our contributions are fourfold: (i) RPKB, a curated knowledge base of 8,191 high-quality R functions; (ii) DARE, the distribution-aware retrieval framework; (iii) RCodingAgent, an R-oriented agent integrating DARE for end-to-end analysis; and (iv) RStatBench, a comprehensive benchmark of 100 statistical tasks. Experiments demonstrate that DARE improves NDCG@10 by 13.43 to 32.65 percentage points across evaluated embedding models, with a MiniLM-based DARE achieving 93.92% NDCG@10 alongside high efficiency. Furthermore, integrating DARE into RCodingAgent nearly doubles mean downstream accuracy on RStatBench, increasing from 31.8% to 62.2% across ten LLMs. This work helps bridge modern LLM automation with R's established statistical computing ecosystem.
RNA molecules often function through multiple experimentally observed conformational states. A central challenge is to capture ensemble diversity without sacrificing structural fidelity: samples may collapse toward averaged representative structures or drift into unsupported conformational regions. We introduce REFLEX (RNA Ensemble generation calibrated by FLEXibility), a sequence-conditioned generator for RNA backbone-frame ensembles based on a heteroscedastic stochastic bridge on SE(3). During training, experimental B-factors calibrate residue-wise bridge widths, exposing each observed conformer through local perturbation neighborhoods with residue-specific scales while keeping constrained regions sharper. A learned flexibility condition guides the bridge-field network to recover target conformer-specific frames from these calibrated noisy states. At inference, REFLEX uses only the input sequence, requiring no MSA-subsampling heuristics. On a multi-conformer RNA benchmark, REFLEX achieves the best precision-recall trade-off for ensemble coverage across evaluation thresholds, while recovery-stress ablations indicate gains in target-frame recovery under stronger perturbations.
George Kontogiannis, Pantelis Tzamalis, Sotiris Nikoletseas
原文摘要
Respiratory sound classification from auscultation recordings is confounded by heterogeneous acquisition conditions, including recording device, auscultation site, and patient characteristics, that introduce systematic variability unrelated to the underlying pathology. Although such metadata is often available, existing approaches either ignore it or rely on naïve fusion strategies that do not model how distinct metadata sources influence different components of the acoustic representation. We propose a family of metadata-conditioning mechanisms for the Audio Spectrogram Transformer (AST), progressing from gated residual fusion to Feature-wise Linear Modulation (FiLM) adapted to the Transformer architecture. We further introduce two novel variants: Token-Aware FiLM (TAFiLM), which applies distinct modulation to classification and patch tokens, and Soft-Factorized FiLM (SoftFiLM), which learns soft masks to align metadata factors with disentangled feature subspaces. Experiments on the ICBHI benchmark demonstrate consistent gains across all conditioning strategies. SoftFiLM achieves 64.11% on the official 4-class task (+3.94% over the baseline) and 72.40% on the 2-class task, establishing a new state-of-the-art on the 2-class task and competitive performance on the 4-class benchmark.
Hiari Pizzini Cavagna, Andrea Proia, Daniele Cesarini, Zeynep Kiziltan, Andrea Bartolini
原文摘要
Generative AI is billed per token, yet tokens are not equally expensive. Each token enters the KV-cache and raises the cost of those that follow, so energy grows non-linearly along a sequence whose tokens look identical. No analytical model states what a token costs, and the most accurate SOTA model is a fit, not a law: it absorbs architecture into fitted coefficients, requiring per-LLM fitting. We derive closed-form expressions of the computational and memory-access complexity of LLM inference in terms of architectural hyperparameters, covering MHA and GQA, and compare a family of analytical models of increasing granularity. From these, we obtain the SIMPLE Inference Model to Predict LLMs Energy (SIMPLE), a four-parameter law that separates prefill from decode, and decode into KV-cache cost and next-token computation. Over traces of 13 dense LLMs from six families on NVIDIA H100 GPUs, it attains R2 = 0.969 and 11.97% MAPE, 3.1-3.5x below the SOTA, without per-LLM calibration. The SIMPLE law explains as well as predicts: prefill energy grows quadratically with the hidden dimension, so larger LLMs suffer most from long inputs; and more aggressive GQA KV grouping reduces the per-token energy by up to 3x relative to standard MHA. Increasing the reasoning token budget pushes output length beyond the analytically derived energy minimum, driving super-linear energy growth; although accuracy grows with the budget, energy grows faster, so Intelligence per Joule peaks at an intermediate budget and falls by up to 3.14x across a 4.0x reasoning-budget range. Agentic serving raises energy per request by 13.0x relative to non-agentic ones, driven by the growth of both the prefill and the KV-cache terms with input length; while prefix caching removes most of the prefill cost (-93%), the KV-cache term is unaffected, so even a cached agentic request costs 4.8x more.
Elizabeth G. Campolongo, Eric R. Sokol, Marty Downs, Gina Glenne, Gregory E. Maurer, Sydne Record, Dustin R. Rubenstein, Matthew J. Thompson 等 12 位作者
原文摘要
Artificial intelligence (AI) technologies and their capabilities have advanced rapidly over the past few years, leading to widespread interest and adoption in the biological and ecological sciences. However, the data infrastructure and computational resources that provide the foundation for AI workflows in ecology and biodiversity science remain fragmented. Moreover, the flow of value remains largely unidirectional: data move from collectors, through repositories and aggregators into models, and finally to research products or management actions. The derived value, such as corrected taxonomies, annotations, quality flags, and insights, rarely returns to source communities. We propose an extension of the original FAIR principles to address the needs of AI-ready infrastructure: the AI-FAIR metadata criteria, which place AI-enabled science in the context of a bidirectional Question-Data-Model triangle. This framework is designed to facilitate large-scale content-based analysis across diverse collections and ensure that the scientists generating the data benefit from the AI-enabled advances built on their data. Alongside these criteria, we introduce an agentic evaluation workflow that both rates existing datasets on the AI-FAIR and traditional FAIR criteria and provides actionable feedback where critical metadata is missing or incomplete. This modular dataset evaluation tool, paired with a domain-appropriate checklist, translates these principles into a reusable assessment of machine-learning readiness, AI-readiness for specified tasks, traceability, and ethical and CARE-aligned stewardship. Through this approach we aim to create a community-informed system to close the model data loop, so that AI also enriches data to further enable scientific discovery.
Ziheng Chen, Jiali Cheng, Zezhong Fan, Hadi Amiri, Diyuan Wu, Gabriele Tolomei, Yang Zhang
原文摘要
Generative recommendation formulates next-item prediction as autoregressive generation over semantic ID (SID) sequences derived from users' historical interactions, making modern recommender systems structurally similar to large language models (LLMs). As privacy and safety concerns grow, these systems increasingly require concept unlearning to remove sensitive or harmful concepts associated with items. However, existing LLM unlearning methods cannot be directly applied to generative recommendation. Unlike word tokens with explicit semantics, SIDs are abstract identifiers that are often shared by both forget and retain items, leading to severe conflicts between concept removal and recommendation utility preservation. To address this challenge, we propose TRACER, an end-to-end concept unlearning framework based on token reassignment. Rather than directly suppressing shared SIDs, TRACER reassigns concept-related items to alternative tokens that better facilitate forgetting while minimizing side effects on retained items. We further introduce a coherence regularizer to preserve semantic consistency among retain items during unlearning. Experiments on real-world recommendation datasets demonstrate that TRACER effectively removes target concepts while substantially better preserving recommendation utility than existing unlearning baselines.
Serdarcan Dilbaz, Arjun Krishna, Kiran Kamble, Daniel M. Bikel
原文摘要
Sparse attention cuts the cost of long-context inference by making the attention matrix sparse: each query attends only a subset of the sequence, chosen by a fixed pattern or, in recent production LLMs, learned per query. In DeepSeek Sparse Attention (DSA), the attention of DeepSeek-V3.2 and the GLM-5 family, every layer is preceded by a lightning indexer that scores the whole cache and keeps the k=2048 highest-scoring tokens for the layer to attend. At long context this indexer scan, not the sparse attention it feeds, is where the decode time goes. But adjacent layers select nearly the same tokens, so a layer can attend the selection of the previous indexer layer instead of its own. This removes the dependency: each indexer then runs concurrently with the attention of its layer. The resulting Multi-layer Indexer (MLI) needs no retraining and no new weights. Its default realization changes no kernel, running inside SGLang's production decode graph; an optional fused decode kernel accelerates it further on DeepSeek-V3.2. MLI composes with IndexCache (the cross-layer reuse GLM-5.2 ships as IndexShare): IndexCache drops the indexer from some layers, and MLI hides the scan on the rest. We serve DeepSeek-V3.2 and GLM-5 under IndexCache's protocol at retentions 1/2 to 1/32 (the fraction of layers that keep an indexer) and contexts 10K to 195K. At equal retention the composition decodes faster than IndexCache: up to 14.5% faster on DeepSeek-V3.2 and 9.8% on GLM-5 at 1/2, and 7.9% and 5.1% at 1/4, largest where IndexCache keeps the most indexers and growing with context. At retention 1/r it is faster still than IndexCache at 1/(2r) while keeping twice the fresh selections: at 128K it reaches 1.32x the unmodified model's decode throughput at 1/4, above IndexCache's 1.28x at 1/8. Its LongBench, RULER and LongBench v2 accuracy stays within a few percentage points of IndexCache's.
Instruction tuning for multi-task network traffic analysis requires data with clearly defined sources, supervision targets, and evaluation conditions. We introduce TrafficInstruct, a newly collected dataset comprising raw packet captures and 55,000 structured supervised fine-tuning records across ten tasks: malware traffic detection, botnet detection, malicious DNS-over-HTTPS detection, Web attack detection, advanced persistent threat detection, encrypted VPN detection, Tor behavior detection, encrypted application/service classification, website fingerprinting, and application classification under concept drift. An automated workflow captures traffic from controlled software executions, attack experiments, and application activities. Labels are supported by task-specific collection evidence, including process or socket attribution, activity records, and execution outcomes. Filtering, sanitization, and task-specific transformations produce instructions, traffic observations, and structured answers. We verify sample sources and partition the dataset by complete collection sessions, keeping all samples from the same session within one partition. The benchmark defines allowed inputs, deterministic label mappings, fixed training/validation/test partitions, and task-wise scoring rules. We evaluate majority-class, random-forest, and ET-BERT baselines, an unadapted 27B language model, and independently trained task-specific LoRA adapters on a shared frozen base. Adapters are selected using validation data and evaluated on the corresponding test sets. Task-specific adaptation improves classification performance, supporting instruction tuning across diverse traffic-analysis tasks.
Gabriel Asher, Cailum M. K. Stienstra, Tomo Oga, Devesh Shah Antonio H. de O. Fonseca, Niall O'Connor, Michael Widrich
原文摘要
A frontier language model was recently reported to solve 1D NMR structural elucidation without tools. We ask whether this capability transfers to tandem mass spectrometry (MS/MS) in the de novo setting, where the model must propose a structure directly from the spectrum without a candidate list or reference library. To separate reasoning from recall, we build a 100-analyte benchmark from MassSpecGym, split by each analyte's textual footprint into in-distribution (well documented in literature) and out-of-distribution (essentially undocumented) arms. We hold the model, harness, and output format fixed and vary only the LLM's access to domain-specific tools. Without tools, a frontier LLM reaches 52% Top-1 accuracy in-distribution but only 2% out-of-distribution. Only access to de novo structure generators improves out-of-distribution accuracy, raising it to as much as 38% Top-1, and the resulting agent outperforms the standalone de novo models both in and out of distribution. These results suggest that LLMs are not effective at de novo MS/MS structural elucidation on their own, and that further gains depend on the specialist tools available to the agent rather than on the LLM itself.
Root cause analysis over telecom and datacenter infrastructures is commonly addressed by methods that consume structured data, from graph traversal and rule-based correlation to probabilistic and learned models. Yet a digital twin of such an infrastructure cannot model everything at a reasonable cost. Operators therefore record much of what changes as unstructured data attached to the graph, where the discriminating evidence often sits. Methods restricted to structured attributes cannot use it. A diagnosis needs a layer that interprets this data. An agentic layer is a natural candidate, since which fragment matters depends on hypotheses that emerge during the investigation. We demonstrate its feasibility with ReAct, the most widely used agentic architecture. A single LLM interprets the descriptions and explores the graph through a typed tool interface exposed via the Model Context Protocol (MCP). On a benchmark of nine synthetic scenarios covering single-node, multi-node and distractor-heavy evidence (360 runs across four LLMs), a frontier-tier model names the root cause and its type on every run. It reaches 96.7% end-to-end diagnostic accuracy, which also requires the exact supporting evidence. A mid-tier model reaches 86.7%. Accuracy drops for every model when the evidence must be combined across several resources.
Hao Fu, Degang Sun, Wan Wei, Jinxia Wei, Chun Long, Peng Gao
原文摘要
Text-to-SQL translates natural language questions into SQL queries, enabling database access without writing SQL. Fine-tuning open-weight large language models (LLMs) is cost-effective, but existing methods treat SQL as ordinary text and optimize token-level likelihood without explicitly modeling its syntactic, semantic, and logical requirements. Thus, a single incorrect alias or comparison operator can invalidate a query while contributing little to training loss. We propose SQLoom, a constraint-aware fine-tuning framework with supervision at three levels: syntactic validity under SQL grammar and database schema, semantic alignment between query decisions and the question, and logical consistency across dependent clause decisions. SQLoom extends supervised fine-tuning with a unified set-contrastive objective favoring sets of correct outputs over minimally different incorrect ones at each level. LLM agents propose this supervision, which a SQL parser and execution on counterexample databases verify. We also release Spider-Refined, a training dataset correcting 378 flawed Spider annotations and enriching database schemas. Experiments on Spider, BIRD, and three perturbed Spider variants show that SQLoom improves execution accuracy for Qwen3-1.7B from 62.9% zero-shot to 88.7% on Spider development, reaching 89.2% on Spider test and 73.4% on BIRD development. It outperforms fine-tuning, reinforcement learning, and agent-based baselines with up to 70 billion parameters. On question perturbations, SQLoom loses at most 0.6 accuracy points versus up to 12.1 for agent-based methods. It generates SQL in a single pass at 0.42s per query, 23$\times$ faster than a GPT-4-based agent method. These results demonstrate the effectiveness of set-contrastive supervision over token-level training.
A memory-augmented agent must decide which memories to keep, merge and retrieve, and its only evidence of their value is its retrieval log: the memories it retrieved in each episode and whether it succeeded. Existing methods assign every memory a value from this log, even when the log cannot separate one memory's contribution from another's. We treat the log as an experiment that nobody designed and ask which memory values it determines. Two counts answer this before any outcome is observed: how often each memory was retrieved, and how often each pair was retrieved together. From these counts, we prove that top-$k$ retrieval leaves every memory value undetermined, however many episodes are logged, and that a question with its own memories, asked once, determines at most one of their values. On fourteen standard memory benchmarks, no memory value is determined in any of 55 configurations. Both causes are choices about retrieval. We introduce LEDGER, which estimates each determined Shapley value in closed form with a confidence interval and reports every other value as undetermined. It merges memories with identical histories without losing a determined value, and it retrieves so that the next log determines more. On pooled HotpotQA, LEDGER raises the fraction of memories determined from 0.001 to 0.797, at a cost of 2.6 accuracy points on HotpotQA and with a gain of 2.7 on LoCoMo.
Ran Ben-Basat, William Kuszmaul, Michael Mitzenmacher, Amit Portnoy, Shay Vargaftik
原文摘要
Uniform random rotations (URRs) are a common preprocessing step in modern quantization approaches used for gradient compression, inference acceleration, KV-cache compression, model weight quantization, and approximate nearest-neighbor search in vector databases. In practice, URRs are often replaced by randomized Hadamard transforms (RHTs), which preserve orthogonality while admitting fast implementations. The remaining issue is the performance for worst-case inputs. With a URR, each coordinate is distributed as a shifted beta distribution, which converges to a Gaussian distribution in high dimensions. Generally, one RHT is not suitable in the worst case, as individual coordinates can be far from these distributions. We show that after composing two RHTs on any $d$-sized input vector, the marginal distribution of every fixed coordinate of the normalized rotated vector is within $O(d^{-1/2})$ of a standard Gaussian both in Kolmogorov distance and in $1$-Wasserstein distance. We then plug these bounds into the analyses of modern compression schemes, namely DRIVE and QUIC-FL, and show that two RHTs achieve performance that asymptotically matches URRs. However, while two RHTs suffice for scalar quantization, they may be insufficient for Vector Quantization (VQ), which depends on the joint distribution across fixed-size blocks of coordinates (as opposed to only marginal distribution convergence for single coordinates). We prove that a composition of three RHTs leads to decaying coordinate covariance. This ensures that any fixed, bounded, multi-dimensional VQ codebook optimized for URRs has the same expected error when using three RHTs, up to an additive term that vanishes with the dimension. Finally, because practical inputs are rarely adversarial, we propose a linear-time $O(d)$ check on the input's moments to dynamically adapt the number of RHTs used at runtime to improve performance.
Matteo Grazioso, Fabio Naccarella, Chiara Gallese, Marco S. Nobile
原文摘要
Artificial Intelligence, and Machine Learning in general, is becoming pervasive across all domains and disciplines. By means of this data-driven approach, algorithms can learns complex task just by learning from samples. However, if the data-set is biased, they might lead to the development of unfair computer systems, that is, characterised by a prejudice towards specific individuals or groups, which might get an unmotivated output from the system. This circumstance lead to the development of multiple metrics designed to quantify the level of unfairness of an algorithm. However, a generalised assumption in the literature is that the sensitive features and the output labels are binary. When the data is not binary, it is forcefully binarized using a one-versus-the-rest simplification. In this work, we discuss the limitations of existing metrics and propose novel, mathematically sound, and generalized multi-group and multi-class formulations of the metrics. Ultimately, our results highlight the empirical assessment distortion caused by binarization strategies, which reveals the limitations of binary fairness metrics in the multi class setting.
Jia Yu, Zilong Wang, Xinyang Jiang, Dongsheng Li, Shuo Wang
原文摘要
Computer-use agents could automate repetitive screen-based clinical work, but their reliability in medical graphical user interfaces remains largely unvalidated. Existing benchmarks focus on general web or desktop tasks and underrepresent medical software, which requires domain knowledge, exhibits markedly different UI design from mainstream applications, lacks public testing environments, and demands safety validation beyond task completion. We introduce MedCUA-Bench, an interactive benchmark for clinical computer-use agents. It covers 18 clinical scenarios across 10 medical domains, reconstructed from real product manuals and open-source medical systems to capture authentic clinical interfaces while avoiding licensing and privacy constraints. Each task ships with paired intent- and step-level goals to disentangle clinical reasoning from UI execution, and is evaluated by a deterministic checker over task completion and five clinical safety dimensions. Across 23 agents, the best closed-source model reaches 54.2% strict success, while all models remain below 9% on the real OpenEMR. Open-source agents average only 2.5%, with the best reaching 16.2%. MedCUA-Bench exposes the gap between current agents and reliable clinical software use, providing a reproducible testbed for future research.