LLM論文ポータル
国内外のLLM・RAG・AIエージェント・AI検索・推論に関する新着論文を検索できます。
OpenAlexから取得できませんでした。取得できた情報源の結果を表示しています。
Role-Based LLM Prompting in Elderly Care Decision Support with Structured Explanations
LLM-Mediated Knowledge Externalization for Low-Resource Sign Language Translation: Open and Closed Expression Knowledge
PlayCraft: Designing Engaging Games from Agent Play Experience
概要を見る
Recent AI agents can generate playable games, but these games are not necessarily engaging. Mastery is a key source of engagement, yet designing for it requires extensive human play and tuning. We propose PlayCraft, an automated framework that uses a vision-based agent to play games and draws on its detailed play experience to revise them for mastery. Starting from an initial game, PlayCraft alternates between playing and crafting. In the playing stage, a player agent watches the game screen, uses the controls, and pursues the game objective over repeated episodes while retaining memory. After each episode, it reports its progress, outcomes, decisions, and reflections. In the crafting stage, a crafter agent uses these reports to revise the game. It estimates player abilities, predicts performance at different proficiency levels, and compares those predictions with a target learning progression. It diagnoses the game and addresses the gaps by giving players meaningful opportunities to improve through better decisions and execution. We evaluate games in fresh runs with player agents across three intelligence settings, measuring learning gain and learning slope. Compared with the initial games, PlayCraft yields larger learning gains and steeper learning slopes in most games. It also outperforms the baselines on the macro-average of each metric at every intelligence setting.
TideCache: Efficient Hybrid LLM Serving with Transparent Cache Management and Layer-Wise Streaming
概要を見る
Hybrid architectures have become mainstream among open-weight language models. Their global-attention and bounded-state layers maintain heterogeneous caches whose demands vary with context length and concurrency. We present TideCache, a serving system on CPU--GPU superchips with three components for allocating GPU memory among caches and managing their residency across GPU and host memory. The Transparent Cache Pool redistributes physical pages among cache components behind fixed virtual addresses, preserving cache layouts and backend interfaces. The Cache Streamer uses computation intervals provided by bounded-state layers to prefetch the next global-attention layer's prefix into GPU slots shared across global-attention layers, reducing GPU residency while overlapping layer-wise transfers with computation. Generated KV remains resident. The Request Scheduler selects resident or streamed execution and triggers capacity adjustments according to request size and queue state. Together, they let HBM released by streaming accommodate other cache components and additional requests. Implemented in SGLang and evaluated on GB300 and GH200 across seven model/platform configurations, TideCache completes up to 2.04x as many agent turns, reduces overall P50 per-token latency by 20--59%, and improves agent P50 TTFT by 1.40--3.05x compared with SGLang with HiCache.
SoK: Assessing the Evolution of LLM Capability in Software Vulnerability Research
概要を見る
LLM-based agents are becoming increasingly capable at discovering, validating and patching vulnerabilities in real software, and vendors and researchers report that this is already changing how vulnerabilities are found and fixed. These claims, however, are scattered across papers, vendor reports and disclosures and are rarely comparable. We introduce a framework that rates these capabilities on an autonomy scale modeled on the levels of driving automation, and qualifies each claim by its correctness, its availability and the strength of its evidence. We use this framework to systematize the published evidence on LLM capability in vulnerability research. We find that agents discovered, validated, and patched vulnerabilities in real-world software without human intervention before the restricted frontier models of 2026, and that open-weight models subsequently attained comparable autonomy for discovery and validation. We then measure the consequences of evolving LLM security capabilities in a decade of public vulnerability records, advisories and exploitation data. In 2026 the number of disclosed vulnerabilities more than doubled, with an increasing share rated critical or high, and severe bugs that had survived a decade or more in widely used code were reported several times more often than before. Yet the time between the stages of a vulnerability's life, from report to fix and from disclosure to attack, has not shifted noticeably so far. We grade the mitigations proposed in the literature with an explicit rubric and draw guidance for each group of stakeholders from the best-supported ones.
Verify the Verifier: Specification-Grounded LLM Cross-Checking for Bugs That Escape UVM Scoreboards
概要を見る
A passing UVM regression shows the design agrees with the reference model on what the scoreboard checks, not that it meets the specification. Bugs that pass every test yet violate the specification, termed scoreboard escapes, survive mature testbenches. We present Verify the Verifier, to our knowledge the first framework whose LLM-based Anomaly Cross-Checker examines a passing regression's traces against the specification. Of 564 bugs inserted into four OpenTitan IPs, 68 are candidate scoreboard escapes: non-equivalent, exercised by the tests, yet passing. The framework confirmed 41 as scoreboard escapes with assertions failing on the buggy design and proven on the original.
Replication of Agentic Security Events: Audit and Investigation (ROGUE-AI)
概要を見る
Agents based on large language models (LLMs) can now delete files and send payments on behalf of their users. News media reports routinely describe failures of these agents (e.g., Replit's coding agent deleting a production database), but a report often only says what went wrong and rarely (in)validates the incident or finds its root cause. Incident catalogs collect these reports, and benchmarks record whether agents fail, but neither replicates an incident nor tests the cause its report might identify. In this paper, we present ROGUE-AI (Replication of Agentic Security Events: Audit and Investigation), a multi-agent framework that closes the above gap and rebuilds each reported incident offline together with a matched control that lacks only the blamed cause, and runs agents on both. Our results suggest that ROGUE-AI is able to reproduce the reported harm in 89% of its 108 attempted replications. Agents built on five current GPT, Claude, Gemini and DeepSeek models made the reported mistake in 18% to 33% of runs, against at most 1% without the blamed cause. Of the 72 incidents with real-world harm, 62 had no underlying attacker, and in these, agents did what their task asked for but missed an ordinary hazard, such as a database shared with the live site. Replications without an attacker caused more harm than replications of prompt injection for all five models. Defenses against prompt injection look for injected instructions, but these replications contain none. Further, a warning after each tool output that names the hazard and its consequence lowered harm for all five models, whereas a note of the same length that does not mention the hazard left most of the harm in place. Built from real incidents, the ROGUE-AI replications form a reusable audit framework that vendors can rerun before each new release.
The Geometry of Lies: Knowledge and Persona Directions for Detecting Deception in Large Language Models
概要を見る
Linear probes are the cheapest LLM lie detectors but often fail to transfer across kinds of lie. Label-free directions from honest and deceptive role-play are untested at 70B scale and on lies about knowledge, where a lie and an honest denial can be identical. We widen a published persona bank with roles that deny knowledge they have, plus honest counterparts. At each model's middle layer the resulting persona direction ranks lies across Liars' Bench better than single-dataset probes on three of its four models; it fails on Qwen, even when Qwen stays in role far more often. On Llama-3.3-70B-Instruct it beats directions from a random persona bank, but its gain over the published direction is within the range of random widenings. On lies about knowledge, persona directions mainly track a reply's wording: where lies and honest denials share their words, our three main contrasts stay inside the random-direction band on Llama and Qwen and reach only its edge on Gemma, though on Mistral one clears it. A knowledge direction contrasts questions the model can answer with ones nobody can, each followed by the same scripted denial. It separates lies from identical honest denials on two models and narrowly on a third, but not on the fourth; its other results partly depend on the benchmark's wording. On the Gender-Secret fine-tune it clears the band by a few hundredths on Llama and fails on Mistral. The sum of the two directions ranks lies better than a supervised probe at the same depth on all four models (on one, both near chance) but is level with the persona direction on Llama's average and behind that probe on held-out lies. More varied role-play questions do not help the persona direction on Liars' Bench.
LLM GerryMandering: Separating Interactions from Additive Effects in LLM Audits
概要を見る
Intersectional bias arises when a model treats people who hold several marginalised identities at once differently from what their separate identities would predict. Auditing such intersectional subgroups remains a challenge: their number grows quadratically, and there is no agreed baseline for what independent attributes would predict. Without such a baseline, many existing audits flag a subgroup when it exceeds each attribute alone, a condition that attributes acting independently already satisfy. We test this standard criterion on 150 synthetic datasets built with no interaction. With the shuffle null, it flags 66.1 false subgroups per dataset on average, and at least one in every dataset. When we plant interactions of known size into these datasets, a repaired null lets the criterion recover only 32% of them. We introduce ADDIS (Additive Intersectional Screen), which fits an additive model over all attributes, scores each subgroup by its Additive Residual Distance (ARD) from that model's prediction, and tests it against a Freedman-Lane null. On our synthetic benchmark, ADDIS flags 0.09 false subgroups per dataset and recovers 84% of the same planted interactions. We then build IndiPersona, 10,000 synthetic Indian personas with six identity attributes, and audit 13 open language models. The standard criterion flags 7.1% of the 738 tested subgroups on average; ADDIS flags 3.0%. Gender identity crossed with sexual orientation is the largest pairing in every model, and counterfactual personas confirm it and additionally reveal a caste x class interaction in all but one model that the observational test lacked the data to detect. Much of what audits report as intersectional bias can thus arise from independent single-attribute effects alone, yet a small core of interactions survives.
Cost-Aware Agentic Feature Optimization System for Fleet-Scale Ranking Models
概要を見る
Large Web platforms operate large fleets of ranking models, each selecting features from a shared pool of tens of thousands. We present an end-to-end agentic system that optimizes features across the fleet under resource constraints shared across models and those set per model, such as feature pipeline cost and serving throughput (QPS), respectively. In this system, LLM agents propose hypotheses, and deterministic tools execute experiments and verify them. For each model, an LLM generates and shortlists cross features from a combinatorial candidate space, and all features are then ranked by several importance methods. Feature blending mixes these rankings into a single selection, using Pareto fronts across tasks for multi-task models. An LLM Researcher agent then tunes the feature selection in 10--25 asynchronous trials under resource caps, and a trimming step keeps the pipeline cost within budget. Across production models, the system improves normalized entropy by 0.31% on average within each model's budget, and the trimming step cuts pipeline cost by 2.3% with no regression beyond measurement noise. The optimized models were launched, and in online A/B tests they improved a primary online metric by 0.20%.
From Seed to BLOOM: Autonomously Growing a Better Out-of-Order Core
概要を見る
Improving microprocessor performance while meeting ASIC implementation constraints like area and frequency is a complex and labor-intensive task that requires expertise across computer architecture, hardware design, and performance analysis and profiling. While large teams in industry can afford the engineering resources to target this problem across performance modeling, RTL design, and ASIC implementations, small teams of researchers and engineers cannot take on this huge engineering burden while keeping up with the pace of advancement in processor microarchitecture. Recent advancements in large language model (LLM) agents present a promising substrate to accelerate this traditionally human-intensive process. In this work, we build and evaluate MiracleGro, an agentic AI-based system that, without human intervention, analyzes cycle-level RTL behavior and synthesis results, identifies performance bottlenecks, and translates architectural ideas into RTL implementation in a leading open-source out-of-order superscalar application core, the Berkeley Out-of-Order Machine (BOOM). Taking into account cycle-exact performance, area, and critical path, MiracleGro automatically produces improved BOOM designs that drastically increase the human-optimized BOOM's performance on substantial workload suites such as SPEC CPU2017 with full reference inputs. MiracleGro achieves an 85% IPC speedup and a frequency improvement of 27% in SkyWater 130nm, resulting in a 2.34x wall-clock time speedup on the full SPEC CPU2017 intspeed benchmarks (30 trillion instructions). This result demonstrates the enormous opportunity to harness frontier models to improve state-of-the-art RTL-driven platforms for implementation-grounded microarchitecture research across hardware designs.
AutoMaster: Autonomous Post-Training by Making the Same Model a Better Teacher
概要を見る
Task success alone does not specify which behaviors post-training should reinforce. Even successful teacher trajectories can contain unsupported assumptions, unproductive repetition, or superficial verification. We propose Skill-FT (Skill-Guided Fine-Tuning), which uses natural-language skills to suppress undesirable habits and elicit underrepresented effective behaviors during teacher rollouts, then fine-tunes a target model on the resulting trajectories. We introduce AutoMaster, an autonomous post-training system in which an LLM agent develops these interventions through trajectory analysis, skill design, data selection, and feedback from rollouts and target-model evaluation. Across Terminal, Search, and software engineering tasks with 4B and 9B target models, AutoMaster improves mean performance over Vanilla SFT on successful unguided teacher trajectories by 1.42-8.99 percentage points, with the teacher, task pool, training set size, and SFT recipe held fixed. Together with the ablations and trajectory analyses, these results suggest that behavioral shaping can make a fixed model a more effective teacher. We further find that retaining skill instructions during SFT yields higher mean performance than removing them across all six settings, despite their absence at evaluation and identical supervised outputs. These findings suggest that preserving skill context as part of a training demonstration can support better behavioral internalization.
JARVIS: An Agent That Designs, Builds, and Tests Circuit Boards on Its Own
概要を見る
Power electronics is central to the energy transition, but developing new hardware still relies on human-led design, fabrication, testing, and iteration. We present JARVIS, an embodied AI agent that autonomously closes this loop for circuit boards. JARVIS maintains machine-specific capabilities and constraints, scans its workspace using a robot-mounted 3D scanner, and constructs an executable digital twin in which it rehearses robot motions, fabrication steps, and circuit behavior before consuming materials. A single text instruction initiates the physical experiment. Using unmodified desktop equipment, the agent coordinates a robot arm, a PCB printer that deposits conductive silver ink and solder paste, and a pick-and-place machine. It verifies progress using cameras and sensors, then tests the fabricated circuit with an OPAL-RT real-time simulator and an oscilloscope. Based on the measurements, JARVIS diagnoses deviations from the design and generates a revised design in KiCad for subsequent rehearsal. In a demonstration involving a 60 Hz notch filter for grid-connected converter sensing, JARVIS completed a first-run experiment in 2 h 39 min without human intervention, achieving 50.6 dB attenuation at 60 Hz. Its digital-twin rehearsal identified five issues before fabrication, including a test criterion that would have incorrectly rejected a functioning board. During execution, the agent also resolved encountered faults autonomously. These results demonstrate an integrated agent-driven workflow for physical design, fabrication, measurement, and iterative improvement of power-electronic circuitry. To our knowledge, JARVIS is the first system in power electronics in which an agent closes the physical design-build-test-learn loop for circuit boards.
Multimodal Error Detection in Chest X-ray Reports: A Large Language Model Reading Image Evidence
概要を見る
Rad-Linter checks chest X-ray reports against image-derived evidence to flag potential errors for review. A pretrained image classifier supplies scores for nine pathologies, and a 1.5B-parameter language model, fine-tuned with LoRA on 1,816 examples, checks their consistency with the report. We compare it with a keyword-and-negation rule baseline given the same reports and scores. Evaluation uses controlled report-score pairings from OpenI, with consistency labels derived from automatic report labels and score thresholds. Each counterfactual quad crosses two real reports with two real score vectors from other patients, creating two consistent and two inconsistent pairs. Every report and score vector occurs in both conditions, so either input alone permits only 50% accuracy. On 131 quads, Rad-Linter achieves 98.5% pair accuracy versus 80.5% for the rule baseline, and classifies all four pairs correctly in 94.4% of quads versus 56.5%. The checkers perform similarly on omission items, including unmentioned and negated findings; larger differences occur on the remaining item types, which require interpreting how findings are described. The same language model without fine-tuning remains near 50% accuracy. On a second set of 181 quads that also includes reports on which the automatic labelers disagree, pair accuracy is 93.4% versus 75.3%. These controlled results show that, given the same image-derived scores, a trained language model detects report-score inconsistencies more accurately than the tested rule baseline, with the advantage concentrated in interpreting how findings are worded.
Held While Waiting: What LLM Agent Sessions Reserve on Serverless and Agent Runtimes
概要を見る
LLM agents spend nearly all their time waiting for the model. On Amazon Bedrock AgentCore, our agents waited a median of 2.7 seconds for each model reply, then computed for an estimated 5 to 40 milliseconds. AgentCore does not charge for CPU during the wait. But it reserves a separate environment, a virtual machine with its own memory, for every conversation, and keeps it reserved until the conversation is closed. So 1,000 conversations reserve 1,000 environments to do the work of 2 to 15 processor cores. A service runs out of environments long before it runs out of processing power. The shortage has two causes, and each has a fix. An environment stays reserved after its conversation ends: in the customer-service benchmark we ran, a conversation lasts 33 seconds on average, but its environment is held for 15 minutes more unless the client closes it. And an environment serves only one conversation, which is mostly waiting: in our simulation, 64 conversations sharing one environment need 59 times fewer environments, and 99 of every 100 replies are delayed by at most 40 milliseconds. The milliseconds are an estimate because they are hard to measure: a stubbed model, a timer around the model call, and steps run back to back each gave a plausible number that was too low. We show how to avoid all three. Of the two fixes, closing needs only the client. Sharing needs the platform, and of the agent runtimes we surveyed only one allows it, in part. They keep conversations apart because an agent runs tools, and sometimes code, chosen by a model. The open problem is to keep them apart without reserving an environment for each while it waits.
A new algorithmic approach for finding flat minima: critically damped Polyak stepsizes
概要を見る
We present a new adaptive stepsize rule that biases gradient descent toward flatter minima in both theory and practice, and establish the first uniform-in-time guarantees for both sharpness and loss reduction for a broad class of locally quadratic objectives. Our critically damped Polyak (CDP) stepsizes multiply the classical Polyak rule by $4$ and a mild damping factor. The factor $4$ is critical in that it produces equioscillations around the manifold of minimizers $\mathcal{M}$ to leading order. Damping provably drives convergence to $\mathcal{M}$. CDP thus operates at the edge of stability by adapting its stepsizes to local curvature. Once the loss is small, its projected iterates follow gradient descent on $\log{λ_1(\nabla^2 \mathcal{L})}$ along $\mathcal{M}$ to leading order, paralleling the flatness bias of normalized gradient descent (Arora et al., 2022). Unlike normalized gradient descent, CDP step lengths shrink with the distance to $\mathcal{M}$, enabling convergence to an interpolating solution that locally minimizes sharpness at finite learning rate. We establish our theory for a broad class of non-negative losses with locally quadratic Morse-Bott structure near $\mathcal{M}$, capturing regular interpolation regimes in smooth overparameterized neural networks. Empirically, CDP stepsizes produce a flattening effect similar to that of Sharpness-Aware Minimization (SAM, Foret et al., 2021), reducing sharpness even in settings beyond our theory, including nonlinear teacher-student networks and fine-tuning of pre-trained LLMs.
A CPU Inside an LLM: Zero-Gradient Compilation of Exact Circuits into Pretrained Weights
概要を見る
Pretrained language models (LMs) remain unreliable on exact multi-digit arithmetic without gradient fine-tuning or external tool interpreters. We introduce closed-form, zero-gradient circuit-grafting primitives that compile deterministic computation directly into pretrained LM weights using zero training examples and zero runtime hooks, while bounding per-layer normalization drift and preserving natural-language (NL) competence. Our construction combines three analytical mechanisms: (1) two-layer Mellin-GLU quadrature over bias-free SwiGLU/GeGLU layers, turning unknown per-token RMSNorm scaling into translation along a log-spaced gate grid ($>250\times$ dynamic range) with IEEE 754 error bounds; (2) cross-token scale cancellation via signed Chebyshev Indefinite-RoPE self-sink attention and memoryless SiLU-antisymmetric GatedDeltaNet shifts; and (3) axis-aligned width extension ($d \to d+K$) with $\sqrt{d/(d+K)}$ norm compensation, Base-1000 canonical BPE emission, and differential $(+254,-256)$ tied unembeddings. An automated 5-stage compiler synthesizes 4-8 digit signed $\{+,-,\times\}$ circuits in constant $\mathsf{TC}^0$ depth alongside 3-literal Boolean (100%, $N=1{,}728$; plus 42-layer Gemma 4 E4B audits) and 5-bit Majority/Parity (100%) circuits. On 6-digit arithmetic ($N=1{,}200$, BF16), IBM Granite 4.2 3B (1.07M-1.33M active non-zeros, $\le 0.036\%$) and Alibaba Qwen 3.5 4B (2.25M-2.39M non-zeros, $\le 0.057\%$) reach 99.17% and 100.00% ALU-prefix accuracy (100% with an FP32 sidecar) and 98%-100% mid-sentence resumption, outperforming 12.8K-sample LoRA (42.2% / 61.7%, 0% inline continuation). A factorized block-sparse sidecar holds A100 memory at 7.89 / 10.65 GiB (4.60 / 8.14 GiB under INT8/INT4 + FP32 sidecar) while passing all 14/14 pre-registered $>2$K-token Tango splits ($N=4{,}392$; 0/11,134 OOD triggers).
Aggron: Aggressive Compression of Agentic KV Caches during Head-of-Line Stalls in the Request Queue
概要を見る
LLM agents run long, multi-turn programs whose contexts grow at every step, and serving them is limited by KV cache capacity. To alleviate this limitation, systems use context compression techniques, which typically wait until the KV cache crosses a large threshold and then summarize large spans of history at once, either in the background while the agent progresses or on the critical path. These summarizations can be done using the same model running on the same engine, or using the same or a different model running on extra hardware. When running on the same engine, summarization competes with the requests being served for both compute and memory capacity. To maximize the cost-effectiveness of summarization, this paper presents a new serving system called {\em Aggron}. Aggron summarizes agentic requests at every turn and selectively summarizes only certain parts of the KV cache. As a result, Aggron retains accuracy and keeps summarization requests lightweight. Moreover, Aggron launches summarization requests from a separate {\em Summarization Queue} while the corresponding agentic request whose KV cache we want to compress is stalled in the main request queue due to head-of-line stalls. The Summarization queue utilizes compute and memory capacity left over by the head-of-line stalls in the main queue. Further, to maximize efficiency, the server issues summarizations only when it predicts that they can complete before the corresponding agentic request will finally execute. In our evaluation, we compare Aggron to state-of-the-art asynchronous summarization. At high load, Aggron reduces P99 program latency by $27$--$31\%$, and raises throughput by $1.34$--$1.40\times$.
MaskPE: One Mask, Two Roles: Weight Placement and MAC Gating in Systolic Arrays
概要を見る
Unstructured pruning reduces weight storage but does not inherently reduce execution cost in conventional weight-stationary systolic arrays. Packed nonzero weights must be restored to fixed processing-element positions, while zero-weight positions must preserve data forwarding. This paper presents MaskPE, a mask-directed systolic architecture that uses positional masks to steer nonzero weights during preload and clock-gate arithmetic registers in zero-weight processing elements during execution, reducing dynamic power while preserving systolic dataflow. At 50% sparsity, MaskPE reduces compressed weight storage by 34.9%, including metadata overhead. RTL validation confirms exact execution with no additional steady-state tile stalls, while gate-level simulations demonstrate 12.5--14.4% lower execution energy across samples of seven LLM projections.
Reasoning-Trace Analysis Using Large Language Models for Interpretable Q-Matrix Construction
概要を見る
Large language models (LLMs) can generate Q-matrices for diagnostic classification models that approximate expert specifications, but evaluations have focused on the final matrices, leaving the basis of individual item-attribute assignments unexamined. This opacity limits the responsible use of AI in measurement. We propose Reasoning-Trace Analysis (RTA), a framework in which an LLM generates attribute-labeled reasoning steps through chain-of-thought (CoT) prompting and these steps are analyzed to reconstruct, compare, and inspect each assignment. RTA is intended to supplement expert review and empirical validation. In a simulation study with perturbed fraction subtraction items, four LLMs recovered the required attributes and withheld attributes the items did not require. In an empirical study with 11 fraction subtraction items, responses from 536 examinees, four LLMs, up to three temperatures, and 100 replications per condition, CoT-derived Q-matrices overlapped substantially with the reference Q-matrix (mean overlap .809) but less than directly generated Q-matrices (.877), and they varied more across replications. Inspection of the traces linked disagreements to item-type coding, labeling rules, and how operations were expressed as steps. Q-matrices with similar overlap differed in fit to examinee responses. RTA provides inspectable evidence that matrix-level agreement and fit do not.
Training Vision Models for LLM-Driven Weight-Space Programming
概要を見る
It has been demonstrated that, in certain settings, large language models can perform weight-space programming to directly modify the weights of neural networks. However, it is unknown how dedicated training of these networks can contribute to the successes of weight-space programming. We present a benchmark of LLM-driven weight-space programming on vision models, in which five LLMs edit convolutional networks and vision transformers on relabeling and error-correction tasks on Waterbirds, CIFAR-100, PACS and Entity-30. We report more than 2,600 such edit processes, whereby each edit is scored on held-out images for success, side effects and accuracy loss. We compare standard training with label smoothing, gradient-based editable training, and our Edit-Ready Training (ERT), a novel method that rehearses the direct, closed-form weight edits LLMs perform. We show that preparing networks for editing raises the success of LLM edits substantially where standard networks are hard to edit. ERT further achieves the highest mean score for GPT-6 Astra and Claude Opus 5.5 across all evaluated baselines. We will release our LLM testing framework alongside all tasks, editing trajectories and scores. Project page: https://viscom.uni-ulm.de/wsp
Democratizing Cross-Project Knowledge for Transforming Water Governance and Climate Resilience: an Agentic AI Ecosystem
概要を見る
EU-funded projects produce extensive evidence in the form of deliverables, reports, and case studies. However, this knowledge is typically scattered across projects and document structures, making it difficult to locate relevant, contextual information when making decisions. This paper introduces an Agentic AI Ecosystem that consolidates heterogeneous project documentation into a domain-aware knowledge base and provides adaptive access through an agentic Retrieval-Augmented Generation pipeline. Documents from three EU-funded projects (IMPETUS, NATALIE, and InnWater) are reconstructed into hierarchical section trees and indexed with structured metadata. Retrieval is supported by two domain-specific embedding models fine-tuned on domain data (water and climate). On top of this knowledge base, a ReAct-style controller iteratively orchestrates retrieval actions by allocating retrieval budget across domain collections and accumulating evidence until a stopping criterion is met. In a retrieval benchmark over the corpus, the proposed approach improves Mean Reciprocal Rank by 26.7% (relative) compared to a single-domain embedding baseline, indicating more consistent retrieval of relevant evidence to support cross-project knowledge synthesis and climate-resilient decision-making.
EgoCaptioner: Towards Comprehensive and Fine-Grained Egocentric Video Captioning
概要を見る
Egocentric video captioning requires fine-grained detail and accurate timing, yet current multimodal large language models (MLLMs) often miss brief actions, confuse their order, or misinterpret the wearer's motion. Existing caption formats also offer limited support for combining detailed narration with precise timing. We present EgoCaptioner, which pairs a video-level overview of the setting, participants, and the wearer's overall activity with timed steps whose narration adapts to varying perceptual demands and marks key moments at specific timestamps. We develop a coarse-to-fine annotation pipeline that establishes event structure before expanding step descriptions, yielding 50,000 training examples from eight egocentric video sources. Complementary rewards assess factual accuracy and coverage, measure step-level temporal alignment, and penalize fragmented or redundant narration, guiding caption refinement through Group Relative Policy Optimization (GRPO). To improve the reliability of factual supervision, we supplement caption references with pose-derived motion facts when sensor data are available. Across five egocentric benchmarks, EgoCaptioner achieves leading performance among evaluated open-source models in caption quality, caption-conditioned question answering, and temporal alignment. It also surpasses Gemini 3.1 Pro on PhysCapBench and all evaluated proprietary baselines on Minerva-Ego under the caption-to-QA protocol. Ablations support the complementary roles of factual and temporal rewards and show that coarse-to-fine annotation provides more effective supervision than single-pass annotation.
HiSA: Sparsifying the Hidden Dimension of RoPE Attention in LLMs
概要を見る
Attention in large language models (LLMs) scores every query against every key with a full dot product. We investigate whether this dot product must be computed completely for attention to be effective. We propose Hidden Sparse Attention (HiSA), which keeps the query and key projections intact but only computes a partial dot product, following a static, parameter-free rule set by the token positions: attention becomes sparse in the hidden (feature) dimension, an axis much less explored than token sparsity, rather than over tokens as in almost all sparse attention methods. Based on the structure of rotary position embeddings (RoPE), we decompose keys and queries into parts which can be omitted when computing attention. By training LLMs from scratch with HiSA, we find that this partial dot product does not hurt but improves performance compared with the complete product while computing only about a third of the query-key products. During inference HiSA improves decoding speed and shrinks memory footprint. We also extend HiSA to FlashAttention-3 kernels for training and inference. Finally, we motivate our approach by showing that dense attention learns to suppress the attention score parts that HiSA drops. Code is available at https://github.com/ambroise-b/hisa.
How Much Should an Agent Remember? Deployment Lessons from Self-Distilled Memory for LLM Agents
概要を見る
LLM agents on long-horizon tasks fail in recurring, procedural ways. They mis-paginate an API, resolve the wrong entity, or answer when no answer was asked. We built a pipeline that lets an agent learn from its own history. It first mines short \emph{lessons} from the agent's own trajectories, one to three per episode. It then consolidates those lessons into a smaller set of \emph{guidelines}, where each guideline records how many lessons it merged. At inference the agent receives either all guidelines or a selected subset. The pipeline ships in the open-source altk-evolve library and is deployed in IBM's CUGA agent and watsonx Orchestrate platform. We report what we learned building, measuring, and productizing it. Our primary instrument is the AppWorld benchmark, across models spanning 40--92\% baseline task completion. The central finding is that the right amount of memory is a property of the model, not the memory. A strong model keeps improving as guidelines are added. DeepSeek-V3.2 gains +9.5 points of task completion and +16.1 points of scenario completion with its full 191-guideline set. A weak model gains less from the same full set than from a small selected subset. gpt-oss-120b starts at 39.9\% task completion. It reaches 51.2\% with all 199 of its guidelines but 55.4\% (mean of two runs) with a selected subset of about 29. Around that result we report what went wrong: a high-support guideline that backfired, provenance silently dropped during consolidation, a swallowed API error that failed every task while producing a plausible log, and a baseline confound that nearly produced a misleading comparison with a prior method. We distill these into recommendations for teams deploying agent memory.
Perfumer: A Large Language Model-based Agentic Framework Accelerating Structural Relaxation and Electronic Band Prediction for First-Principle Calculation
概要を見る
First-principles structural relaxation and band calculations require substantial computation and expertise. Here we present Perfumer, a large language model-based agentic framework that couples Quantum ESPRESSO with two predictive models. Across tested structures, pre-relaxation with Perfumer-relax model followed by QE refinement yielded a 2.75-fold overall speedup over QE alone. Perfumer-band model combines a periodic graph neural network with a latent Hermitian spectral head returns 12 band energies at any requested wavevector, preserving reciprocal lattice periodicity and time-reversal symmetry without constructing Hamiltonian. Trained on 125,153 self-curated band structures, Perfumer-band model reaches a median per-material mean absolute error of 0.157 eV on 2,607 test materials, including lanthanide, actinide and magnetic compounds, from held-out composition families spanning 85 elements, taking about 97 ms per material on two consumer CPU cores. The agent turns natural-language requests into relaxation-to-band workflows, from a single uploaded structure to automatically constructed new candidates and returns predicted properties.
Training LLMs to Follow Reasoning Instructions with Minimal Edits
概要を見る
Large language models increasingly rely on extended reasoning traces, yet the training signals used to shape this behavior remain coarse. Supervised fine-tuning on full trajectories forces the model to imitate an entire demonstration, overconstraining its output distribution, while trajectory-level reinforcement learning provides only weak credit assignment, rewarding or penalizing whole completions without identifying where they succeed or fail. We propose edit-based training, a general framework for learning from a model's own trajectories via localized textual corrections. Given a sampled completion, an editor identifies the first point of failure under a specified criterion and proposes a minimal replacement span. The model is then supervised only on these edited spans. By design, this concentrates supervision where the model's reasoning first breaks down and aims to preserve its existing reasoning patterns elsewhere. The method applies both offline and online, with the latter continually adapting supervision to the model's evolving failure modes. We instantiate this approach on reasoning instruction following with Qwen3-32B on the ReasonIF benchmark. Compared to full-trajectory imitation and trajectory-level reinforcement learning baselines, edit-based training substantially improves compliance with the reasoning instruction while largely preserving the model's skills on untargeted tasks. These results suggest that localized correction is an effective complement to these approaches for shaping how LLMs reason.
Sirius-Tiamat: How Small Can a One-Pass Decision Encoder Get Before Its Decisions Change?
概要を見る
A one-pass decision encoder reads a text, a question and several named options, and returns a probability for each option in a single forward pass. We ask how far Laya, a 421M-parameter encoder of this kind, can be compressed before its decisions change. We count decisions that differ from Laya's own full-precision (fp32) answer, and track calibration (how well confidence matches accuracy, measured by expected calibration error, ECE), on 1,221 synthetic items with the training schemas, 273 with an unseen schema, and 800 out-of-family questions from ARC, MMLU and HellaSwag. Layer pruning of a related model family is known; we add flips, calibration, held-out data and retraining. bf16 changes 0.2% of in-family decisions but 2.2% out of family; 6-bit grouped rounding 1.3% and 5.9%; deleting 6 layers 0.9% and 9.1%, while ECE rises from 5.9% to 39.2%. One fitted temperature, needing no labels, restores it to 3.6%. One scale per tensor fails below six bits, and CPU dynamic int8 changes 30% of decisions, nearly all through one layer. Total variation tracks flips for rounding but not for deletion, and an early exit saves layers over a static cut only if its confidence gate is first recalibrated. Distilling a 16-layer student cuts its in-family flips from 44% to 9% (45% out of family). At equal stored size, rounding flips fewer decisions than deletion. Data are synthetic and LLM-labelled, for one architecture.
One Framework to Diagnose Them All: A Stratified Benchmark for Bengali LLM Hallucination
概要を見る
Hallucination research in low-resource languages fragments easily: each newly named failure mode prompts its own dataset, spending scarce annotators on incomparable resources. We argue for consolidation and test it at community scale. Alikbachan is a Bengali hallucination-detection benchmark stratified along three crossed axes: a human-validated cultural-distance band, six task types, and a five-family, 32-mode hallucination taxonomy. It comprises a harvested 2,516-item test set (Phase 1) and a quota-constructed, gold-labeled 5,000-item set (Phase 2). The stratification is diagnostic. Four zero-shot LLM judges vary by over 30 macro-F1 points across strata, and the item-level predictions of 99 Phase-1 and 32 Phase-2 competition systems reproduce the profile: discourse-level omission defeats judges and most teams alike, whereas only a minority over-flag negative controls. The judges' verdict probabilities rank items well (ROC-AUC 0.74-0.95) but are overconfident (ECE 0.04-0.29). An open competition on Alikbachan drew 222 teams and 3,566 submissions, and a single-prompt 32B judge reaches the 57th percentile. Of the 25 teams in both phases, 24 regress on Phase 2 (mean $-17.6$ points) while every zero-shot judge improves, and Phase-1 rank barely predicts Phase-2 rank (Spearman $ρ=0.35$), so a Phase-1 leaderboard alone would have crowned a different winner.
ClosureBench: A Benchmark for Reasoning under Closure Assumptions
概要を見る
AI agents built on large language models (LLMs) often answer questions from incomplete records. Missing information might be interpreted as unknown or false, depending on the completeness of the source. Current evaluations fail to distinguish these two aspects, as they are usually interwoven with other reasoning demands. We present ClosureBench, a benchmark that tests whether LLMs follow the stated rules for dealing with missing information. Every base scenario is tested under OWA, CWA, and LCWA, and only the treatment of missing information changes between variants. ClosureBench provides a verified base benchmark as well as agent extensions for action choice, source scope, and updates. A model is considered correct on a base scenario only if it provides correct answers in all of its closure variants. Across 960 instances, overall accuracy can mask substantial closure mistakes. Models achieve overall accuracy from 71.18% to 99.86% across replicated runs, but their core switch accuracy, measured on the scenarios whose gold label actually changes with the contract, varies between 18.23% and 99.31%. Large discrepancies emerge under LCWA. For example, some agents choose unknown despite the prompt indicating that missing information should be treated as false. Our findings demonstrate that evaluations of AI agents should not only measure whether agents can reason over existing facts but also whether they use the appropriate assumption when information is absent.
OceanCoder: A Verification-Guided Large Language Model for Automated Simulation Modelling in Ocean Engineering
概要を見る
Numerical simulation is important in ocean engineering, but model construction requires expert knowledge. Large language models support automated modelling, but code syntax and successful execution alone do not establish whether generated models and results meet engineering requirements. We present OceanCoder, a Qwen3-8B-based large language model for automated simulation modelling in ocean engineering. To address specialised programming requirements and scarce training data, we develop an extensible data construction pipeline producing 12,000 verified supervised fine-tuning (SFT) examples across 12 engineering task categories. A three-layer verification framework checks model inputs against engineering parameters and task constraints, verifies completion of required simulation analyses, and assesses the engineering plausibility of computed outputs. Its outcomes provide graded rewards for group relative policy optimisation (GRPO), guiding the model towards engineering task completion. We also construct the first benchmark for automated modelling in ocean engineering, covering out-of-distribution tasks and real-world engineering cases. Experiments and ablation studies demonstrate the effectiveness of the data pipeline, verification framework and reward design. OceanCoder outperforms both base models and the closed-source comparison, with Pass@1 increasing from 0% for the base model to 46.77% after SFT and 51.43% after GRPO. These results support engineering verification as a training signal for automated simulation modelling.
Rigel-Ladon: An Option-Count Law for a One-Pass Decision Encoder, Tested on Held-Out Option Counts and Held-Out Schemas
概要を見る
Laya is a small neural network that reads a text, a question and K written options, and returns one probability per option in a single pass. We ask how its accuracy falls as K grows from 2 to 35 and whether that can be predicted. We score nested option sets (every K-option question contains the smaller ones) on four question pools, for the released model (Base) and our fine-tune (FT), and score every item against each single rival. FT trained on one 12-option topic question (in-family); a 12-option HR (department-routing) question is out-of-family. Logit accuracy (the log of correct over incorrect answers) is linear in log(K-1) with slope 0.68 to 0.88. Fitted on K up to 6, the law predicts K=7 to 12 within 0.6 points and K up to 24 within 1.2 (a log-linear rule: 1.6 and 6.7). Taking the slope from the topic question and measuring only two-option accuracy on HR predicts its other K within 0.3 points (FT) and 0.7 (Base). On a 35-label pool of five schemas the law fails (5 to 11 points). Independence of irrelevant alternatives is rejected: the gold-versus-rival log-odds shifts by +0.24 to +2.07 nats per doubling of K, yet pairwise answers reproduce K-way accuracy within 4.7 points. Order averaging (averaging over several orderings of the options), shortlists, a temperature law and calibrated prediction sets did not transfer (FT's 90% sets cover 84% on HR). 2 of 10 predictions made in advance held. Caveats: one model family and LLM-labelled synthetic text.
取得元:OpenAlex / arXiv / J-STAGE。CiNii Researchは連携準備中です。
Powered by J-STAGE · J-STAGE WebAPI利用規約
J-STAGEは記事の公開日、その他は論文の発行・投稿日で絞り込みます。日付のみのデータは日単位で判定します。取得範囲には上限があり、すべての関連論文を網羅するものではありません。
Thank you to arXiv for use of its open access interoperability.