Pretraining Exposure Explains Popularity Judgments in Large Language Models
Analysis of popularity bias in OLMo LLMs traced to pretraining exposure in Dolma corpus via entity-level statistics.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Analysis of popularity bias in OLMo LLMs traced to pretraining exposure in Dolma corpus via entity-level statistics.
Adaptive policy optimization removing hyperparameter tuning for RL post-training under distribution mismatch.
DRIFT: discrete flow matching for offline-to-online RL using CTMC policies with advantage weighting.
ProfiliTable: multi-agent framework using profiling-driven agentic workflows for table cleaning and transformation.
LLM agent framework for post-hoc crop yield forecast correction using domain tools on strawberry and corn data.
GAP framework fixes feature-space mismatch in multimodal LLM visual reasoning by aligning latent token generation with input embedding norms.
Study shows passage convergence—how effectively hints eliminate wrong answers—improves LLM performance on inferential QA over retrieved answers.
MetaColloc meta-learns neural basis functions offline to solve PDEs at test time without retraining, replacing optimization with collocation assembly.
The feature is designed to help people get real-time context about trends and breaking stories, as well as receive recommendations, all within conversations.
Frontier models (Opus 4.6, GPT 5.4, Gemini 3.1) miss dangerous coding agent actions 2–30× more often after 800K tokens, exposing context-length monitoring gaps.
QAP-Router frames NP-hard qubit routing as dynamic quadratic assignment, using RL to exploit logical-qubit interactions for quantum compilation.
Extends SAGA agentic AI governance framework to decentralized Byzantine-resilient setting, protecting against malicious distributed providers.
Adapts differential evolution optimization to quaternion-valued search spaces, potentially improving model compactness and AI training efficiency.
MedHopQA benchmark evaluates LLM biomedical reasoning via multi-hop disease-centered questions, resisting answer-elimination and training data contamination.
Linearized Graph Sequence Models reframe graph message-passing as sequence modeling to decouple computational depth from propagation depth.
δ-mem adds compact online associative memory to LLM backbones via delta-rule updates, enabling efficient long-context reuse in agentic systems.
The family of a 19-year-old college student is suing OpenAI over claims that his conversations with ChatGPT led to an accidental overdose. In the lawsuit filed on Tuesday, Sam Nelson's parents allege ChatGPT "encouraged" the teen to "consume a combination of substances that any…
Community member shares video tutorial and GitHub repo implementing a minimal Claude-like coding assistant from scratch.
GPT-5.5 achieves first solve on ProgramBench hard/extreme tasks, substantially outperforming Claude Opus 4.7 on novel SWE benchmark.
Open-source TUI tool provides visibility into Claude Code agent loops, costs, and security issues; author reports $14K spend, 20% wasted iterations, and 3 credential leaks over 90 days.
Community milestone: 1M datasets published on Hugging Face, celebrated as progress for open-source AI.
OpenAI CEO Sam Altman has begun his testimony against Elon Musk in a high-profile jury trial in a California federal courtroom. Altman, alongside OpenAI president Greg Brockman, is a primary defendant in the trial brought by Musk. Altman, Brockman, and Musk were all part of the…
Hollywood actors and producers are standing behind a new AI licensing standard that will tell AI systems whether they'll need to pay to use a person's likeness, creative work, characters, and designs. With the Human Consent Standard, people can set terms for the use of their…
Rivian's AI-powered voice assistant is rolling out today to the company's vehicle fleet. The assistant will be available through a software update to all compatible Rivian Gen 1 and Gen 2 vehicle owners who subscribe to the company's Connect Plus cellular service, which costs…
So I thought I was assigning Claude a simple task of uploading a hero image from a folder within Cowork to a Wordpress page using the MCP connector. First it tried to compress the hell out of the image... nobody asked it to do that, then it went through all kinds of failed…
OpenAI case study on Codex adoption for finance workflows: MBRs, reporting, variance analysis, and planning scenarios.
Reddit discussion of an unverified prompt guidance guide claiming to optimize Claude Opus 4.7 behavior; anecdotal user experiences without empirical validation.
Developer demonstrates local coding workflow using Qwen2.5-Coder-7B autocomplete and Qwen3.6-35B agentic model on RTX 5080 with RAM offloading.
Isomorphic Labs closes $2.1B Series B led by Temasek and Loweringthe Bar to scale AI for drug discovery and protein folding.
Link: techcrunch.com