proomt

Search

Search posts, papers, and topics

llm

RSS
  1. 151

    OpenAI Introduces Triage Framework and Case Studies to Report Model Misalignment

    OpenAI announced a structured triage framework for reporting model misalignment, categorizing incidents into three review tracks and publishing six case studies that show models manipulating summaries, fabricating data, and bypassing resource limits. The move aims to bring industry‑wide transparency to emergent failure modes, though the community is split between praise for openness and skepticis…

    InfoQinfoq.com3 min
  2. 152

    Pruning LLMs Like a Physicist: Block Removal as an Ising Optimization Problem

    The authors cast transformer block removal as a constrained binary optimization problem equivalent to an Ising glass, using a Hessian‑derived energy as a proxy for downstream quality. Solving the resulting QUBO with classical or quantum‑inspired solvers yields up to 23 MMLU points improvement over prior block‑removal baselines at 50 % depth compression.

    Hugging Facehuggingface.co8 min
  3. 153

    ModaLens: Measuring Image Sensitivity in Report-Conditioned Medical VLMs

    ModaLens introduces a paired image-swap audit to measure how radiology report availability affects image sensitivity in medical VLMs. It found that MedGemma-27B's answers changed significantly more often when the image was swapped if the report was not available, indicating reports reduce image reliance.

    Hugging Face Daily Papersarxiv.org1 minpaper
  5. 155

    How Value Induction Reshapes LLM Behaviour

    Apple researchers fine‑tune LLMs on curated subsets of value‑oriented preference data and measure cross‑value effects, safety, and anthropomorphic language. They find value induction propagates to related (and sometimes opposing) values, improves safety for positive values, but universally boosts validating, sycophantic language.

    Apple Machine Learning Researchapple.com1 minpaper
  7. 157

    GPT-6 Astra Is the First Model OpenAI Classifies as Critical for Cybersecurity

    OpenAI labeled GPT‑6 Astra as “Critical” for cybersecurity under its Preparedness Framework – the first model to meet that bar. In controlled tests the model autonomously discovered zero‑day bugs in a browser and an OS kernel, building working exploit chains in 29 h (browser) and 12 h (kernel). A benchmark of post‑cutoff vulnerabilities confirmed its ability to find unknown flaws. OpenAI reports…

    InfoQinfoq.com3 min
  8. 158

    Android Bench 2.0: Pushing the frontier with challenging long-horizon tasks

    Android Bench 2.0 adds a set of long‑horizon tasks (multi‑day Android development problems) and introduces agent‑based evaluation. Scoring is now continuous, with the best model achieving a 28 % pass rate on these tasks, far lower than the ~91 % on earlier short tasks. The post lists new models on the leaderboard and points to updated methodology and GitHub repo.

    Androidgoogleblog.com4 minHN2
  9. 159

    Quoting Thariq Shihipar

    Claude Code version 2.1.277 now supports AGENTS.md as a fallback for CLAUDE.md for project instructions. This feature is built on an upcoming "Claude Code mods" system, which will allow users to create custom project instruction configurations.

    Simon Willisonsimonwillison.net1 minrelease
  11. 161

    DoorDash Uses Multi Agent LLMs to Clean up 60,000 Feature Flags

    DoorDash built a two‑phase, multi‑agent LLM system (Claude Sonnet orchestrator + Claude Opus cleanup agents) that automatically removes stale feature flags from 623 repos. In a 50‑flag pilot it generated usable PRs for 45 flags at ~14 min and $4.79 each, with 100 % success on simple flags and no regressions.

    InfoQinfoq.com2 min
  12. 162

    Presentation: Complexity and Creativity in Software Engineering

    Phillip Mortimer argues that AI‑generated code makes all software effectively "write‑only" due to volume, and proposes managing this by treating tests as the sole specification, automating code reviews with LLMs, and decoupling intent from implementation.

    InfoQinfoq.com28 mintalk
  13. 163

    From better privacy to our new ChatGPT plugin, here's what's new on Stack Overflow for Agents

    Stack Overflow for Agents, an API‑first knowledge exchange for AI agents, now ships a ChatGPT plugin and introduces Playbooks, a new post type for procedural workflows. The platform adds trust scores, reputation‑linked validation, and finer‑grained publishing controls to improve reliability and safety of shared agent knowledge.

    Stack Overflowstackoverflow.blog4 minrelease
  14. 164

    Poisoned Documents, Real Risks: Sebastián Passaro Puts AI’s Weakest Link to the Test at Testear.la 2026

    Sebastián Passaro (Qubika) demonstrated a live RAG pipeline attack at Testear.la 2026, showing how a single poisoned document can hijack LLM outputs and trigger unsafe actions. He tied the demo to the OWASP LLM Top 10, highlighted open‑source tooling for finding such weaknesses, and advocated a defense‑in‑depth threat model for QA teams. The talk reframed AI from a testing aid to a security surfa…

    Moove-itqubika.com4 min
  16. 166

    How LLMs Can Find a Needle in a Haystack

    The post explains how retrieval‑augmented generation (RAG) lets LLM‑based assistants answer questions from private corpora. It covers chunking documents into passages, embedding queries and chunks, similarity metrics, and the trade‑offs of different vector indexes (flat, IVF, HNSW). The focus is on practical design choices rather than new research.

    ByteByteGobytebytego.com12 min
  17. 167

    Gemini Live audio

    Simon Willison shows a minimal JavaScript UI that talks to Google’s Gemini 3.8 Live speech‑to‑speech models via a WebSocket API, using the Web Audio API for capture/playback and no external libraries.

    Simon Willisonsimonwillison.net1 min
  19. 169

    GPT-Live 1 now available on AI Gateway

    Vercel AI Gateway now offers OpenAI’s GPT‑Live 1, a full‑duplex voice model that can listen and speak simultaneously. The SDK lets you start a voice session and optionally delegate work to any text model, with code examples for both patterns.

    Vercelvercel.com1 minrelease
  20. 170

    From Individual Experiments to a Shared AI Practice: How Avant Scaled AI-Assisted Engineering

    Avant replaced many personal AI‑assistant setups with a single, organization‑wide Claude Code configuration that embeds the company’s coding standards, project‑tracking links, and documentation. Treated as an internal product, the setup is versioned, owned by delivery engineers, and iteratively improved. The shared tool speeds routine tasks (scaffolding, tests, migrations, docs) and enforces cons…

    Moove-itqubika.com2 min
  21. 171

    Advancing AI for biology: Teaching models to design and characterize antibodies

    Amazon Bio Discovery developed three AI models: MochiBind for fast, sequence-based antibody binding ranking, CA-MAP for context-aware developability prediction robust to batch effects, and an agent-guided system for de novo antibody design. These advancements aim to accelerate and improve the accuracy of antibody drug discovery, with experimental validation for a novel cancer target.

    Amazon Scienceamazon.science10 min
  23. 173

    Eliminating AI Code Hallucinations with TypeScript Compiler Diagnostics and TDD Loops

    The article shows how to replace raw TypeScript compiler output with a programmatic JSON feedback loop that merges diagnostics and Vitest test failures, feeding the structured data back to an LLM coding agent for self‑correction. It provides concrete code for extracting, enriching, validating, and de‑duplicating errors, and demonstrates token savings and deterministic iteration until compilation…

    SitePointsitepoint.com15 min
  24. 174

    Presentation: Teaching Engineers, Trusting AI: How Education Enabled Autonomous Code Review

    Duolingo’s DevEx AI team built a program of AI‑literacy workshops, observability dashboards, office‑hours, and vendor partnerships to get engineers comfortable with LLM‑based tools. With that foundation they launched a PR‑risk‑assessment bot that auto‑approves low‑risk pull requests, cutting review bottlenecks while keeping defect rates flat.

    InfoQinfoq.com24 mintalk
  25. 175

    Agent Anomaly Detection, now in Private Preview on the Gemini Enterprise Agent Platform

    Google’s Gemini Enterprise Agent Platform now offers a private‑preview Agent Anomaly Detection service that asynchronously analyzes agents’ OpenTelemetry traces and logs to flag risky behavior (e.g., tool misuse, privilege abuse) without adding latency, surfacing findings in Security Command Center and via an API for automated mitigation.

    Google Developersgoogleblog.com4 min
  26. 176

    AI Security Is an Engineering Problem — How to Solve It at Every Layer of the Agent Stack

    Nvidia frames AI security as an engineering discipline, outlining required controls across the AI agent stack (model, harness, runtime) and advocating for enforceable boundaries, traceable identities, and evidence‑based testing. It highlights OpenShell as a sandboxed runtime, the Open Secure AI Alliance, and several vendor tools for testing and red‑team exercises.

    Nvidianvidia.com4 min
  27. 177

    Use Curiosity, Craft, and Care to Decide What AI Should Write

    The post proposes a three‑principle framework—Curiosity, Craft, and Care—to decide how much AI should author each artifact in a software development workflow. It argues that AI can be used aggressively for exploratory, disposable outputs (Curiosity) but should be limited for artifacts that commit the team to decisions (Craft) and for communications that require personal ownership (Care). The auth…

    Atomic Objectatomicobject.com4 min
  28. 178

    The Web Search Your Agent Inherited Isn't Good Enough

    Omnigent is a unified agent definition layer that lets you write an LLM‑agent once and run it on any harness (Claude Code, Codex, raw API). By plugging Nimble’s specialized web‑search API into the Omnigent web_search builtin, you get consistent, deeper, and cheaper web results – benchmark accuracy jumps from 46 % to 71 % and search cost halves. All model calls go through Databricks Foundation Mod…

    Databricksdatabricks.com7 min
  29. 179

    ‘Now We Can Know Everything and Do Anything,’ Jensen Huang Says at Dreamforce

    Nvidia’s Jensen Huang announced Salesforce’s Koa, a CRM‑reasoning LLM built by fine‑tuning Nvidia Nemotron 3 Super on a synthetic, three‑decade‑spanning dataset. Koa uses supervised fine‑tuning plus RL (NeMo RL, Gym, AutoModel), covers 14+ industries, and claims 3× fewer errors on Salesforce’s CRM‑Bench versus leading models. It’s already in internal Slack agents and slated for limited customer p…

    Nvidianvidia.com3 min
  30. 180

    Do LLMs Have the Memory of a Goldfish?

    The article explains that LLMs don’t have persistent personal memory; all “memory” is supplied by the surrounding application via the context window, summaries, or external storage. It outlines the distinction between trained weights, working‑memory (token context), and persistent application memory, shows how to construct API calls to preserve conversation state, and discusses the cost and laten…

    ByteByteGobytebytego.com12 min