Researchers propose Decentralized Language Models (DeLM), a coordination layer that lets multiple LLM agents work asynchronously through a shared context and a task queue with no main agent in the loo…
JetBrains released Mellum2.1, an Apache 2.0 licensed 12B mixture-of-experts thinking model with 2.5B active parameters, aimed at coding agents. Reinforcement learning in real repositories raised the m…
A new arXiv paper (2610.02351v1) introduces DeReAct, a modular agent architecture that externalizes action validation and completion control into two gating policies: a Critic that validates proposed …
Google Research open-sourced EnvHarness on October 2, a programmable wrapper layer that adapts existing agent training benchmarks by modifying starting conditions, action contracts, and task sequencin…
Researchers from KAIST and Seoul National University identified why retrieval-based speculative decoding loses its speedup in coding agent pipelines: the suffix index stores repository files as raw on…
Google Cloud AI Research released RRSI, an open-source framework for Regularized Recursive Self-Improvement of Agent Harnesses, on September 21, 2026, in collaboration with UNC-Chapel Hill, Stanford, …
Researchers released TomasuLLM, a runtime that executes LLM agent tool calls out of trajectory order while preserving task-execution correctness, according to arXiv paper 2609.38201v1. TomasuLLM draft…
A paper posted September 2, 2026, "EarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction" (Shi, Sun, Dong, Wan, Lo, Gu), proposes halting agent benchmark runs whose outcome is already predi…
Researchers submitted TokenCast, a method that learns a composable cost representation for each execution segment of an LLM agent run, forecasting token consumption before and during execution without…
Google Cloud AI Research, with UNC-Chapel Hill, Stanford and Washington University in St. Louis, released RRSI (Regularized Recursive Self-Improvement), an Apache 2.0 research framework that lets an L…
A Peking University paper (arXiv:2609.31430) by Zhensheng Zou, Guoqing Wang, and Dan Hao introduces Latent Observations, Hard Actions (LOHA) and Anchored Context Distillation (ACD), a context-compress…
A developer's analysis of roughly 4,300 real coding-agent sessions (Claude Code and Codex, 43 developers, ~350,000 LLM steps) found that the median LLM step re-sends about 119,000 cached prefix tokens…
OrcaRouter released OrcaSAQ2 27B, a 3-bit mixed-precision quantization of Qwen3.8-27B that shrinks the checkpoint from 54GB in BF16 to 12.3GB, a 77.2% reduction, under Apache-2.0. The model card repor…
A developer's analysis of coding agent architecture shows that swapping only the harness around a fixed model — keeping weights, tasks and context window constant — raised SWE-bench Verified bug-fixin…
A head-to-head comparison finds Anthropic's Claude Sonnet 4.5 outperforms OpenAI's GPT-5 on coding and agentic benchmarks, scoring 77.2% on SWE-bench Verified versus GPT-5's 74.9%, and 50.0% versus 43…
Epoch AI launched Benchmark Reviews on September 17th and classified nine of its first 15 reviewed benchmark versions as Flawed, four as Verified, and two as Not Enough Info, flagging defects in tests…
Researchers introduced block parallelism (BP) and context-sharded block parallelism (CSBP), new distributed parallelism dimensions for training block diffusion language models (BDLMs) on long contexts…
Mutagent Helix published a meta-evaluation comparing its agent-building tool against hand-driving Claude Code across twenty agent benchmarks, measuring each calibration pass by dollar cost and number …
Salesforce AI Research published DarwinX, a framework that evolves AI agent harnesses without modifying model weights, raising pass@1 on the 1,260-task WebArena-Infinity benchmark from 43.5% to 93.0% …
Nexpath, a middleware layer that appends verification steps, constraints, and acceptance criteria to AI coding prompts, improved Claude Code's performance on SWE-bench Verified from 27 of 40 tasks sol…