Large Language Model News

Researchers developed GAVEL, an LLM judge protocol that compares clinical timelines against their source case reports, returning a discrepancy type, verdict, and report passage for each difference. Evaluating 2,738 findi…

A new web-agent architecture called OdoBot completes web-application tasks using 44% and 80% fewer tokens than the state-of-the-art agents Agent-E and WebVoyager, respectively, according to an arXiv paper (arXiv:2609.134…

AutoTailor, a meta-agentic framework from researchers behind arXiv paper 2609.13548v1, cut an initial 1,283 unrefined web-agent APIs down to 87 through offline filtering and then to a 33-API set via Dynamic Reselection. …

A carbon-aware routing framework that distributes function-calling queries across a three-tier edge-cloud architecture reduces operational carbon emissions by 4x on average while matching cloud-level accuracy, according …

A new arXiv paper (arXiv:2609.13805v1) proposes an LLM-enhanced multi-agent reinforcement learning framework that simultaneously optimizes the grid, electric vehicles, and charging stations within a unified loop for publ…

Researchers presented SAILOR, a proof-of-concept system that detects unsupported numerical choices in natural-language optimization problems, asks users targeted follow-up questions, and updates the optimization model be…

Researchers introduced ClinAgent, a conversational agentic Retrieval-Augmented Generation (RAG) system built on a Large Language Model agent following the ReAct paradigm, which lets clinicians and researchers query clini…

A first-move retrieval module called Question's Gambit raised downstream agent answer accuracy from 83.1% to 90.5% with gpt-5.5 over the Pi-Serini agentic baseline on the BrowseComp-Plus benchmark, according to the arXiv…

A new arXiv paper (2609.13995v1) argues that the debate over synthetic data in marketing research should shift from whether LLM-generated respondents work to when they work, distinguishing three types of synthetic data: …

Researchers introduced UniCAR-RL, an annotation-free reinforcement learning framework that decouples perception and reasoning optimization in Multimodal Large Language Models (MLLMs) to improve complex mathematical visua…

A new guide on optimizing AI agent cost, speed, and quality recommends routing individual agent steps among GPT-4o-mini, GPT-4o, and GPT-5.6 Luna based on task complexity, tool requirements, confidence scores, and evalua…

Researchers submitted a paper to arXiv on 14 Sep 2026 introducing Discrete Beckmann Transport Models (DBTM), a time-independent flow whose autonomous transport map provably carries any point in the ambient space to a fix…

A June 2026 working paper from Ramp economists Ara Kharazian and Ryan Stevens with Revelio Labs' Lisa Simon, pairing Ramp spend data with Revelio Labs payroll records across more than 21,000 US businesses, found that hea…

A research team has released ZGCM-1, a fully open 7B dense foundation model trained from scratch for math and agentic search, built on the premise that compact models cannot passively memorize the open web but can overco…

Bolt.new launched a mode called Forge on September 14, 2026 that gives every individual Pro subscriber up to 50 times more usage on four open-weight coding models — GLM 5.3, GLM 5.3 Flash, Kimi K3, and DeepSeek v4 Pro — …

Shipd.ai, operated by Datacurve, runs a contributor platform where engineers from 40+ countries solve STEM problems — software engineering, machine learning, and data science — to produce training data for frontier LLMs.…

A proposed method called exploration-guided prompt scaffolding aims to fix the equal rollout budget that standard online reinforcement learning gives to training prompts, which vary widely in how informative they are for…

A developer building an AI technical interviewer replaced LLM-driven conversation flow with a finite state machine after the model repeatedly skipped stages, ended interviews early, and mishandled transitions roughly 20%…

A developer has designed a deterministic AI agent execution runtime in TypeScript that replaces the naive LLM agent loop with a Finite State Machine architecture built around four phases: Observe, Propose, Verify, and Co…

A calculator built by Sunk Cost estimates that a $3,499 Apple M5 Max Mac Studio with 64 GB of unified memory running Alibaba's quantized Qwen3.8 27B model would need 44 years to break even against hosted inference for a …