The Archive

Paired-residual subspace removal improves multilingual text embeddings from frozen LLMs by disentangling language variation.

DLite enables history-free speculative decoding for agentic LLM inference, recovering 80% degraded acceptance rates in long-context workloads.

Structured dataset of 5,346 Italian laws (1985–2026) extracted via LLM from parliamentary records, covering procedural events.

Mental World Model framework for simulating subjective cognition and Theory of Mind extends LLMs beyond linguistic inference with dynamic evolution.

LLM medical verification fails to distinguish independent trials from same-trial republication, revealing aggregation brittleness across five model conditions.

RoBERTa-RT model measures cognitive load of LLM-generated vs. human text via reading time but findings are psycholinguistic not AI-architectural.

VISTA infers latent user values from survey responses and uses hypernetworks to steer frozen LLMs toward personalized outputs without prompt engineering.

HardEval benchmark measures LLM/agent capability to implement STIG security controls on live systems with persistent state changes.

Framework assessing reliability of LLM agents simulating stakeholders for multi-criteria healthcare operations decision-making under preference uncertainty.

EEG decoder analysis across Llama-3.1-8B layers reveals passage-level context structure in silent reading aligns with deep layer embeddings.

PresenceBench benchmark measures LLM adaptation to user functional state (readiness, flow, stability) via receptivity and specificity metrics.

LLMs trained on experimental feedback learn to propose model improvements without expensive RL, enabling recursive self-improvement for AI research automation.

SchemaShield-Lite repairs LLM agent function-calling robustness against interface drift via schema validation and retry logic without model retraining.

AgentOnto guides LLM agents on enterprise analytics using ontology-structured knowledge and iterative feedback incorporation across databases and documents.

SRCoT introduces a closed-loop Retrieve–Reflect–Revise pipeline for edge LLMs to generate evidence-grounded text with provenance tracking.

Token probability scoring method infers native language background from non-native English text using Llama 3.1 on three datasets.

MobileInject reveals cross-modal injection vulnerabilities in MLLM-based mobile agents through adversarial multimodal content embedding.

Enterprise data assistant deployment study shows knowledge graph-based semantic mediation improves LLM data access for imperfect business user prompts.

Analysis and fix for token repetition failures in diffusion-based multimodal LLMs using information flow tracing and cache optimization.

Annotrace: tool for structured step-level annotation of LLM agent execution traces across multiple trace formats.

JointCore method and JointCoreBench benchmark for interaction-aware memory selection in LLM agents, identifying jointly necessary evidence.

FLARE: physics-grounded benchmark for evaluating LLM agents on electrical grid feeder restoration with incomplete observations across 78,750 episodes.

Systematic evaluation of confidence-based small-to-large LLM routing across 72 configurations and 42K predictions shows mixed multilingual reliability.

RIVET enables turn-level routing in multi-turn LLM interactions by using post-call response evidence during training while keeping deployment decisions pre-call.

MultiAgent-OPD combines on-policy distillation with RL to jointly train hierarchical multi-agent systems (Leader-Searcher-Answerer) end-to-end using LLM judge feedback.

DIA shows test-time compute from single models provides correlated rather than independent evidence; proposes disagreement-triggered independent arbitration across models.

Evaluation framework validating LLM-generated social networks against survey data from ANES 2020-2022 study.

On-policy bidirectional distillation strengthens MoE LLMs by leveraging internal expert capacity as teacher signal.

Method detecting LLM hallucinations by measuring context reliance via dark-space projection of hidden states.

Anthropic demonstrates Claude optimizing quantum circuit design, showing LLM application to specialized physics and hardware problems.