Researchers introduced AFDBench, the first benchmark for evaluating generative meteorological reasoning, comprising 7,732 expert-written Area Forecast Discussions from 13 National Weather Service offi…
Researchers introduced LACL-GUI, a length-aware contrastive learning framework for GUI agents powered by multimodal large language models, which improves reinforcement learning stability and performan…
Researchers developed a memory-efficient, single-GPU training and inference architecture for autoformalizing natural language mathematics into Lean 4 statement code, using the Qwen3-Coder-30B-A3B spar…
Amazon Web Services (AWS) published a technical guide on designing custom reward functions for multi-turn reinforcement learning (RL) with Amazon Nova Forge, emphasizing that the reward function deter…
Researchers introduced Task-Aware Prompt Rewriter (TAPR), a model that reformulates user prompts into task-optimized prompts to improve downstream LLM performance, trained via reinforcement learning w…
Researchers present B1ade, an efficient retrieval-augmented generation (RAG) architecture with a 335M parameter embedding model and a 1B parameter small language model (SLM). B1ade-embed achieves top …
A new study, 'The Physics of Multi-Turn Long-Horizon Planning,' introduces a controlled environment to isolate variables in agentic planning. The research identifies three stages—pre-training, post-tr…
A new two-phase pipeline called HeuristicEdu aligns Qwen2.5-7B toward Socratic tutoring, improving Scaffolding Effectiveness from 30.0% to 63.3% and reducing keyword leakage from 30.0% to 13.3% on hel…
The llm-d project introduces co-operative time-slicing to interleave independent reinforcement learning jobs on shared hardware, increasing aggregate accelerator duty cycles from ~40% to 70% without i…
Researchers Anna Manaseryan and Casey Kennington presented an online reinforcement learning framework at the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue that adapts rob…
A research team has demonstrated that a compact Small Language Model (SLM) retrained for control reasoning and embedded in a validator-guided correction loop can achieve 91.5% average action-alignment…
Researchers introduced Rank-Then-Act (RTA), a framework that trains AI using video-derived ordinal signals instead of traditional rewards, achieving state-of-the-art results on benchmarks like PyBoy a…
Researchers introduced PPRO (Profile-guided Personalized Retrieval Optimization), a framework that enhances memory retrieval for long-term conversational agents by integrating user profiles. The syste…
Researchers introduced TREK (Teacher-Routed Exploration via Forward KL), a new AI training method that enhances learning through unconventional exploration strategies. TREK significantly improved perf…
Researchers at an undisclosed institution applied Group Relative Policy Optimization (GRPO) to adapt an LLM-based automatic speech recognition (ASR) system to synthetic speech, achieving a 40% relativ…
Researchers introduced BV-Blend, a critic-free reinforcement learning framework that stabilizes advantage estimation for aligning large language models by blending prompt-local on-policy statistics wi…
Researchers have developed Agentic Monte Carlo (AMC), a method that enables reinforcement learning-style optimization of black-box LLM agents without requiring parameter-level access. The technique us…