LLMs
Latest breakthroughs in Large Language Models
Articles
OpenDiscoveryTrace: A New Benchmark for Auditing the Reasoning⭐8
OpenDiscoveryTrace introduces a 558-trajectory benchmark for auditing AI scientists' reasoning, revealing process-level failures invisible to output-only evalua...
When Agent Governance Helps⭐7
Governed autotelic AI agents under guardrails: GAMPO framework tests show governance benefits depend on model capacity and domain, with minimal interventions of...
How Much Does Corpus Choice Change Dependency-Distance Estimates?⭐10
Comparing 38 parallel UD treebanks reveals that mean dependency distance is corpus-dependent: ranking reversals exceed sampling error, yet dependency-length min...
EXAONE Forecast for Finance: A Linear-Time Foundation Model for⭐9
EXAONE Finance: a linear-time foundation model for financial time series forecasting, outperforming state-of-the-art methods on FinVerse benchmark.
Counterexamples as Feedback for Agent Self-Correction⭐9
A-CEGIS framework uses counterexample feedback to evaluate multi-turn regex synthesis, achieving 90% task resolution within four turns and surpassing single-sho...
Computer Science > Computation and Language⭐9
Study compares bounded personas vs. retrieval for a frozen language agent, finding retrieval matches on classification tasks but falls short on regression.
Speculative Macro Commit: Accelerating Tool-Using LLM Agents⭐8
Speculative Macro Commit accelerates tool-using LLM agents by predicting multi-step actions, cutting latency by up to 44.9% while matching accuracy.
Governing Enterprise Analytics: A Policy-Controlled Approach to⭐9
A governed framework for AI-driven analytics separates intent interpretation from policy-controlled SQL execution, improving replayable query reliability.
Where Does Harness-Oimization Value Live? Localized Gains and⭐8
Nearly all optimization value in self-evolving LLM agents resides in reflection/control slots, not role or strategy, as revealed by HARNESSEVO.
AI-Driven Structure and Implementation of Next-Generation⭐7
AI-powered adaptive English textbooks improve learner outcomes through personalized tasks, real-time feedback, and teacher analytics, boosting accuracy and enga...