Tag
Kimi K3, an open model, rivals Fable in agentic tasks, and routing between them achieves 93% accuracy with up to 50x cost savings, establishing a new state-of-the-art approach.
Kimi K3, a new AI model with 2.8 trillion parameters and 1 million context length, has been released on web and app, featuring leading capabilities in coding, agentic tasks, reasoning, vision, and agent swarms.
This paper proposes ReOPD, a method for on-policy distillation of LLM agents that reuses pre-collected teacher trajectories as replayed prefixes, achieving improved efficiency and accuracy without new environment interactions.
Proposes SEED, a self-evolving on-policy distillation framework that converts completed trajectories into hindsight skills to improve reinforcement learning for interactive agent tasks, achieving consistent performance gains and sample efficiency.
LLM-as-a-Verifier is a simple, cheap, general-purpose self-improvement technique for agentic tasks, using fine-grained scoring and logprob-based ranking to achieve SOTA on multiple benchmarks like SWE-Bench Verified and Terminal-Bench V2.
This paper from Stanford, NVIDIA, and UC Berkeley introduces LLM-as-a-Verifier, a training-free verification framework that uses continuous scoring from LLM logits to improve accuracy across coding, robotics, and medical domains, achieving state-of-the-art results on multiple benchmarks.
This paper from Stanford, Berkeley, and NVIDIA introduces LLM-as-a-Verifier, a general-purpose verification framework that uses token logits for continuous scoring. It achieves SOTA on multiple benchmarks including Terminal-Bench V2 (86.5%) and SWE-Bench Verified (78.2%), and provides fine-grained signals that can accelerate RL training.
Introduces Object-Centric Environment Modeling (OCM), a method that organizes LLM agent experience into two executable code bases (object knowledge and procedure knowledge) to improve reuse, verification, and reduce invalid actions in interactive environments.
TREK is a staged procedure that uses distillation to expand exploration support for policy optimization, improving performance on mathematical reasoning and agentic tasks beyond standard GRPO.
A review of the new Ornith-1.0 35B coding model that bypasses public benchmarks and tests it on real agentic tasks, highlighting its strengths in long-horizon coding and coherence, as well as costs like verbosity.
Discusses why GLM-5.2 moved away from GRPO, suggesting that GRPO's assumptions may not hold for long-horizon agentic tasks.
Counsel is the first public dataset of human meta-evaluations of LLM critiques for agentic tasks, designed to improve the calibration and reliability of automated evaluation methods.
This paper shows that a carefully crafted data recipe for long-context reinforcement learning, using minimal outcome-based GRPO, significantly improves reasoning across multiple models and benchmarks, and transfers to agentic tasks like GAIA and BrowseComp.
Proposes SelSkill, a dual-granularity preference-learning framework that learns when to invoke skills in agentic tasks, improving task success by 10.9% on ALFWorld and 5.7% on BFCL.
Maxime Labonne shares that their model is trending on Hugging Face and is surprisingly capable at agentic tasks despite having only 1B active parameters.
Mimo V2.5 offers performance comparable to Claude Opus 4.5 at a fraction of the cost, making it a highly cost-effective AI model for agentic tasks.
Forge is a reliability layer for self-hosted LLM tool-calling that uses guardrails and context management to dramatically improve performance on multi-step agentic tasks, lifting an 8B local model from 53% to 99% accuracy.
Noah Ziems expresses excitement about their recent work in Pedagogical RL, which aims to transform data collection for complex agentic tasks like coding.
OpenAI releases GPT-5 in their API platform, a state-of-the-art model achieving 74.9% on SWE-bench Verified and excelling at coding, agentic tasks, and long-context reasoning. The release includes three model sizes (gpt-5, gpt-5-mini, gpt-5-nano) and new API features like verbosity control, minimal reasoning mode, and custom tools.