The Archive

Simple LLM harnesses outperform hand-crafted RCA agents on microservice root cause analysis via RCAtlas benchmark.

Pairwise Bellman Validation enables tuning-free model selection for off-policy evaluation without auxiliary regression.

SEMA-Auto enables autonomous multi-turn jailbreak attacks on LLMs by reading responses and adapting termination in real time.

Statistical framework for LLM evaluation via confidence-stratified sampling and LLM-as-judge auditing with calibration.

EvoRouter: self-evolving framework optimizes multi-agent LLM routing via insight-guided on-policy distillation from sparse rewards.

Domain-specific LLM framework extracts structured data from Pashupatinath Temple reviews using entity taxonomy.

Hybrid framework extracts causal chains from 56,202 aviation reports via LLM and knowledge graph for physics-based failure quantification.

Philosophical critique arguing fixed-benchmark evaluation cannot establish LLM semantic competence without dynamic context.

Fine-tuning on goal-goal conflicts improves LLM compliance to system policies over conflicting user instructions.

Question-Driven Evidence Reasoning Architecture improves LLM in-context learning by surfacing implicit rubrics beyond stated user requests.

Conflict of Skills Attack exposes security gaps in composable LLM agent systems where isolated skill validation misses downstream policy violations.

ComFair benchmark audits intersectional fairness across 479k dual/multi-attribute LLM instances, exposing harms missed by single-attribute evaluation.

RL method trains LLMs to abstain with user-specified confidence targets, enabling inference-time accuracy-abstention tradeoff adaptation.

PatentAgentBench evaluates LLM agents on realistic patent workflows spanning technical, legal, and linguistic tasks without reference materials.

BellowsKV shares KV-cache across transformer layers on demand, reducing memory for long-horizon LLM agents with minimal accuracy loss.

Multiscale Policy Optimization aligns credit assignment to turn-level for long-horizon LLM agents, improving search and tool-use tasks.

CoCHI improves LLM calibration by identifying and intervening on internal heads responsible for overconfident errors via causal analysis.

Dual-specialist framework with LLM judge balances evidence-grounded reasoning and reference-matching for traffic anomaly explanation.

LLM agent extracts GLP-1 medication experiences from Reddit using knowledge graphs for pharmacovigilance analysis.

Study shows LLM target models excel at self-critique but fail at generating novel red-team attacks.

C-SafeGen provides certified safe LLM decoding via claim-based streaming guardrails with formal jailbreak guarantees.

BriLLM-MetaPred implements unified representation-prediction framework for language modeling based on AGI theory.

Benchmark and open-source engine for reversible pseudonymisation of Russian/English documents sent to external LLMs, evaluating five tools.

Five LLMs execute harmful tool calls 27–99% when given misinformation; credibility judgment is unreliable across models.

Continual Learning Exam benchmark with 120 expert-authored tasks measuring long-horizon LLM agent improvement in professional workflows.

Theory-grounded taxonomy of model curiosity in LLM interactions, finding curiosity questions are 17–28% of all questions.

Taxonomy and deterministic classifier for Pydantic schema-conformance failures across nine LLMs beyond aggregate validity rates.

WebViSE extracts visual design patterns from web to enhance LLM website generation from natural-language specifications.

Counterfactual Test-Time Augmentation removes spurious correlations in LLMs without retraining by generating counterfactual edits via auxiliary models.

11k controversial-topic dataset with value-based personas reveals safety gaps in LLM alignment when persona-conditioned requests span diverse value systems.