LLMs
Latest breakthroughs in Large Language Models
Articles
HakemBench: A Turkish Benchmark of Typed Decisions for LLM EvaluationNEWβ8
HakemBench introduces a Turkish benchmark of 2,346 items for evaluating LLM probabilistic decisions across seven tracks using calibration-aware metrics and boot...
MintFlow: Training-Free Minimal Trajectory Intervention forNEWβ8
MintFlow introduces a training-free constrained sampling method that minimally perturbs pretrained flow matching trajectories, enforcing constraints without ret...
The First Token Is Not the Verdict: Hidden Costs of Reading LLMβ8
Reading LLM judge verdicts from the first token overstates position bias by up to 42 points, per new researchβa hidden cost for anyone auditing judge reliabilit...
SCM-Based Fairness and Faithful Explainability for Legalβ8
SCM-based debiasing of LegalBERT fails to reduce gender disparity yet consistently degrades SHAP explanation faithfulness across five seeds, revealing a dissoci...
From Proposal to Verified Effect: Praxa, an Evidence-Boundβ9
Praxa presents an evidence-bound harness for governed AI agent execution, with explicit state tracking, audit results, and pilot benchmarks from proposal to ver...
When Do Causal World Models Help Modular LLM Agentsβ8
Causal world models improve modular LLM agents most in structured tool environments where cross-module interfaces are identifiable and decision-relevant at acti...
On-Device Named-Entity Recognition: A Deployability Study ofβ9
A deployability study of on-device named-entity recognition across nine systems, evaluating accuracy, latency, output validity, and the trustworthiness of model...
Heavy-Tailed Memory Traces in Long-Horizon Language Agentsβ9
Long-horizon language agents show heavy-tailed memory traces. This study introduces CTWM, a rank-based memory controller that cuts tail prediction error and tok...
Improving OCR Faithfulness via Gated and Attenuated On-Policyβ8
GAD-RL adaptively gates and attenuates teacher supervision during on-policy distillation, improving OCR faithfulness and achieving strong CHAOS-Bench and OmniDo...
FD-VAD: Semantic Endpoint Detection for Streaming Full-Duplex Speechβ9
FD-VAD introduces an ASR-free streaming endpointer that makes semantic Continue/Stop decisions directly from causal audio windows, achieving top EOT recall on T...