The Archive

Multi-expert LLM framework using bipolar argumentation to extract MITRE ATT&CK tactics from cybersecurity intelligence reports.

ReGap-Bench identifies binding failures in multimodal LLM reasoning and evaluates visual scaffolding fixes.

Functional invariances enable low-precision quantization for linear recurrence LLM architectures during inference.

Self-evolving agent audits SWE-bench dataset quality, test coverage, and specification issues for LLM coding benchmarks.

Empirical study of policy judgment robustness in Claude Opus, Gemini Flash, and Mistral on 40 presidential actions and 15 hypothetical reforms.

OPERA is an LLM agent that designs CMOS opamps end-to-end from specifications to layout using gm/ID method and open-source EDA.

Formulaic uses LLM agents to co-optimize hardware-software AI infrastructure design across boundaries, verified by mathematical solvers.

AgentDecept is an executable benchmark of 477 scenarios measuring deceptive behavior in tool-using LLM agents under pressure in stateful workflows.

Claude Opus 5 infers difficulty patterns in USMLE/PANCE exam items using LLM-induced rubrics to identify interpretable factors beyond prediction.

LLMs interpret domain semantics to guide simulation-based engineering redesign, reducing computational cost of optimization loops.

SocialArena benchmark evaluates LLM social reasoning across mental health, tutoring, and companionship via three-tier framework.

IMRouter routes temporal patterns in LLM-based time series forecasting to mitigate frequency imbalance and rare-pattern errors.

Ted Sanders argues model size explains most performance variation but no universal task-domain ordering exists across LLM capabilities.

Sleep framework enables LLMs to retain knowledge via episodic memory replay during continual learning.

Keepace.ai, domain-specialized LLM based on Qwen3.5-27B for Chinese-language fitness guidance with curated corpora and expert benchmarks.

IFAAB benchmark audits GPT-4o financial advice disparities across gender, caste, education, and geography in Indian personas.

3DCodeEdit benchmark evaluates LLM instruction-guided 3D Blender program editing across 2,592 tasks.

TIDE optimizes FSDP2 collective communication during LLM training by eliminating layout conversion overhead in gradient aggregation.

Rubricon automatically generates and refines evaluation rubrics for LLM assessment via discriminative sampling without manual annotation.

Multi-agent clinical LLM systems amplify demographic disparities in triage regardless of model or aggregation on MIMIC-IV data.

GAAC achieves lossless KV cache compression via exponent-redundancy coding for long-context LLM inference on edge devices.

Paper identifies Role Grounding Misattribution vulnerability in LLM agent memory systems via social interaction attacks.

Skill-Uni: unified RL framework jointly optimizing skill retrieval and utilization for LLM task-solving.

Automated workflow converts natural language to power electronics schematics using LLMs with constraint verification.

DecompGym benchmarks LLM agents on algorithmic decomposition tasks (Benders, column generation) across 81 instances.

Collaborative LLM decoding leaks private context via message transcripts despite public-only API views; includes leakage audits.

LLMON post-training framework uses structured notation to reduce prompt injection and improve control, trained on 4.6B tokens.

Epistemological analysis applies justified-true-belief framework to distinguish when LLM outputs constitute rational knowledge.

TGTV framework integrates spatiotemporal graph learning with LLMs for interpretable UAV fault detection and diagnosis.

Critical evaluation reveals CLadder benchmark's procedural-generation artifacts may artificially inflate LLM causal-reasoning scores.