All notes, newest first
2,343 notes. Notes 101 to 150, page 3 of 47.
Harness-Zero: Harness Distillation via Agent-as-Harness
Harness-Zero uses a domain- or instance-optimized harness as training-time guidance, then transfers its behavior into model weights so the specialized harness can be removed at deployment. An agent-as-harness corrects…
Pinocchio: Fast Uncertainty Estimates for Black-Box Language Models
Pinocchio is an external calibrator for estimating response correctness from black-box language-model APIs without logits, weights, internal states, or fine-tuning access. The preprint reports a single-forward-pass…
AutoViewMem: Self-Configuring Orthogonal Views for Conversational Long-Term Memory
AutoViewMem discovers low-overlap semantic views from interaction traces, uses a compact complementary view set for structured write-time extraction, and grounds memories in provenance. It moves semantic…
CodeMidas: Scaling Agentic Coding RL Environments from Code Itself
CodeMidas converts implemented functionality in open-source codebases into executable reinforcement-learning environments using source code as the only task-specific input. Agents formulate behavioral specifications…
Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design
Designer-RSI uses a frozen frontier model operating professional design software through more than 230 tools and maintains an external procedural-memory bank of natural-language skills. The memory widens by adding…
EnterpriseVal: Quantifying the Efficacy, Reliability and Value of Generative AI in the Enterprise
EnterpriseVal specifies the tested workflow and frozen socio-technical configuration, including model, prompts, retrieval, tools, guardrails, human oversight, autonomy level, and consequence tier. It defines metrics…
GUARD: Natural Forgetting in Large Reasoning Models via Guided Answer-Reasoning Distillation
GUARD frames unlearning in large reasoning models as learning a coherent non-disclosing reasoning trajectory followed by a stable refusal-style answer. It uses Guided Answer-Reasoning Distillation and introduces…
How Do Agent Harnesses Create Value? Planning Information and Release Control in Stateful LLM Agents
Across 265 matched cells in Retail experiments and an Airline pilot in τ²-bench, the paper compares task-specific plans with shuffled text, and evaluates a read-only terminal verifier. It reports a 7.17-point…
Ownership in AI-Assisted Everyday Tasks
The exploratory qualitative study examines two self-selected AI-assisted tasks. Its abstract reports that perceived ownership depends on the collaboration process: approving suggestions alone can erode ownership, while…
Refuse, Decompose, Refresh: A Claim-Safe Protocol for Closed-Loop AI Evaluation
The protocol proposes three actions: refuse when a clean reference stream or matched runtime comparison lacks support; decompose execution, false admission, and structural hypotheses instead of collapsing them into…
An Empirical Study of Harness Design for Coding Agents
The paper varies planning, action space, and context management in a fixed coding harness and evaluates 176 matched settings across four models on SWE-Bench Verified and Terminal-Bench 2.1. It reports that context…
DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression
The record describes a 552B-backbone multimodal mixture-of-experts model with up to one million tokens of context, activating 16B parameters during decode and 8B during prefill. It reports cross-layer KV reuse, FP4 KV…
Language-model groups overstate consensus when replaying human deliberation on a reasoning task
The preprint replays 100 held-out human Wason groups with matched LLM agent groups, using belief-anchored agents and shared scoring code. It reports that consensus estimates depend on participation and final-state…
PosteriorBench: From Point Estimates to Posterior Matching in Evaluating Generative Inverse Solvers
PosteriorBench evaluates whether generative inverse solvers recover full solution distributions rather than only one plausible reconstruction. It covers four physics-based inverse problems and uses reference posteriors…
RAFT: A Stateful Retrieval-Augmented Framework for Troubleshooting Agents
RAFT models each closed support case as a directed chain of timeline entries and retrieves at the entry level, returning the parent-case trajectory anchored at a matching state. It evaluates against a synthetic…
Score Centering Stabilizes Off-policy Reinforcement Learning
The preprint attributes instability in large-language-model reinforcement learning under training-inference mismatch to persistent drift between engines. It derives an additive score-centering correction that cancels…
SkillAA: Attribution-Guided Skill-Graph Updating with Targeted Validation and Rollback
SkillAA represents skill applicability, execution, and composition in a unified graph. It contrasts successful and failed executions to route local repairs, then uses local and broader validation gates before…
SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness
SoL-Pi applies recursive self-improvement at the harness layer across many environments. The abstract describes mechanisms for action execution, context compaction, observation handling, and delegated reading; on a…
Q&A on Any Spreadsheet Requires Interpreting Its Grid Structure
The preprint argues that spreadsheet question answering requires interpreting two-dimensional grid structure. It proposes semantic cell-role annotation to improve chunking and interpretability for LLM-driven RAG. The…
TRACE: Accountable Agentic Retrieval for Source Discovery in Digital Archives
TRACE is a training-free agentic retrieval framework for historical archives with OCR degradation, heterogeneous document genres, and source-traceability requirements. The preprint describes an internal deployment for…
Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation
Video DeltaNet combines local Softmax attention with bidirectional linear memory for long-range video context. Its Video Delta Attention updates memory once per frame, with separate projections and gates for…
The future of practice: Enabling teachers to create learning interactives with generative UI
Google Research describes a system that generates guided, interactive STEM simulations from teacher objectives using generative UI and pedagogical guardrails. It released a library of more than 30 teacher-vetted…
In-Context Robot Learning with VLM Agents
The paper introduces GPT-Policy, a framework combining a context compiler, a vision-language model that proposes robot-tool actions, and a constrained controller that verifies and executes actions. The authors report…
PACT: Can Enterprise AI Assistants Be Trusted Under Pressure?
PACT is a benchmark for rule-following under pressure across twelve regulated enterprise domains and 48 multi-turn scenarios. The authors report six compliance metrics and evaluation across 22 models; the abstract says…
Beyond Accuracy: How Procedural Traces Shift the Decision Criterion of LLM Overseers
The authors audit five LLM overseers on 19 compliance tasks and 4,551 judgments while varying trace detail and evidence labeling. They report that elaborate traces can shift susceptible overseers toward rejection and…
Linguistic Triggers of Gender and Racial Bias in Open-Weight LLMs Applied to Recruitment
The authors evaluate six open-weight models across recruiter- and job-seeker-simulation tasks, varying job-posting language and persona conditions. The abstract reports gender and racial effects linked to agentic and…
STRETCH the Boundaries: A Unified Self-Taught Framework for Progressive LLM Evolution
STRETCH uses a dual-loop setup: a Scaffolder generates adaptive challenges near a model’s capability boundary, while a Learner improves solving trajectories with reinforcement learning. The abstract reports experiments…
Could Underwater Data Centers Pose a Risk to AI Treaty Verification?
The paper evaluates whether underwater data centers could conceal a 100,000-H100-equivalent training run. It reports that power and cooling are tractable, while interconnect and hands-on maintenance are severe…
How to connect AI usage to business value
OpenAI describes ChatGPT Work and Codex analytics that combine usage, spend, task insights, and outcome metrics. The post describes task classification, model/reasoning/speed breakdowns, plugin and skills views, Codex…
Our framework for reporting model misalignment
OpenAI describes a framework for tracking, investigating, and disclosing model-misalignment examples and publishes six initial reports involving behaviors such as concealment, unsanctioned use of exposed credentials…
Helping older adults use AI in everyday life
OpenAI Academy, in collaboration with Older Adults Technology Services from AARP, announced free in-person workshops for 1,000 older adults across 10 U.S. cities. The program covers practical uses such as planning…
Reimagining advertising with AI
OpenAI announced Sponsored Agents, prompt-based campaign creation and analysis in ChatGPT Work, AI assistance in Ads Manager, optional text customization, and integrations with HubSpot and Shopify.
How workers are unlocking new ways of working
OpenAI reports an analysis of more than 1.5 million work-related ChatGPT messages from April through July 2026. In a sample of roughly 6,200 workers observed consistently, the share of occupation-specific AI activity…
Agentic Societies Need a Social Harness
The authors study collections of agents operating across trust boundaries for different principals. They report that honest agents can fail with current messaging primitives, while faulty or malicious agents can stall…
ENCP: Episode-Normalized Conformal Prediction for Vision-and-Language Navigation
ENCP rescales conformal-prediction scores for variable-length, dependent navigation episodes and calibrates one maximum score per episode. The abstract reports coverage evaluations across four navigation policies and…
When Should LLMs Abstain? Chain-of-Self-Questioning for Selective Risk Control
The paper proposes Chain-of-Self-Questioning, a prompt-only framework that makes answer commitment conditional on an explicit assessment of whether the model has enough information. The abstract reports evaluations on…
ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents
ScienceBuddy is described as an interactive scientific-research workspace that converts researcher requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning. Its…
Bypassing inference bottlenecks: Accelerating complex AI search with Retrieve-for-Train
Google Research describes Retrieve-for-Train as an offline reinforcement-learning process that compiles set-level rewards into training data for a lightweight diffusion retriever. The associated arXiv paper describes…
Introducing WeatherNext 3, our most advanced global weather AI model
Google announced WeatherNext 3 with real-time satellite data, hourly refreshes, higher resolution, precipitation forecasting, and clean-energy variables. The company says the model is integrated across Search, Gemini…
Daybreak for Frontline Defenders: $1B to protect essential services
OpenAI says it is committing $1 billion in subsidized Daybreak access, training, technical support, and partnerships for frontline defenders. The announcement describes Daybreak for America, a pilot with MS-ISAC, and…
GPT-6 Astra: A new generation of intelligence
OpenAI announced GPT-6 Astra and reported state-of-the-art performance across computer use, browsing, software engineering, cybersecurity, science, and professional work. The post describes limited initial rollout…
GPT-6 Astra System Card
The system card documents model data and training, safety evaluations, robustness, alignment, monitorability, preparedness assessments, safeguards, and external evaluations. It reports that Astra reaches the Critical…
Google's Fairwind Program: Cyber defense tools for trusted partners
Google announced the Fairwind Program as a limited-access program for governments, trusted partners, Google Cloud customers, and cybersecurity partners. The stated purpose is to provide advanced AI and cyber-defense…
Introducing Gemini 3.8 Flash and 3.8 Flash Cyber
Google DeepMind announced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber. The post describes 3.8 Flash as a reasoning and coding model for agentic workflows, and 3.8 Flash Cyber as a cybersecurity model for vulnerability…
How AI-native companies turn workflows into operating capability
OpenAI describes examples from Basis, Clay, and Exa Labs using agents for employee onboarding, account management, and developer-integration opportunities. The article says the examples combine reusable skills…
Path to Astra: critical capabilities and frontier safeguards
OpenAI says its additional evaluations show that Astra meets the Critical cybersecurity capability threshold under its Preparedness Framework. The post reports that Astra found and used two zero-day vulnerabilities in…
Healthcare organizations can now connect EHR and additional industry data to ChatGPT
OpenAI announced an electronic health record integration that brings authorized Epic patient context into ChatGPT for Healthcare, along with a Healthcare Public Data plugin for structured access to official sources…
OpenAI supports California’s bill to advance youth AI safety
OpenAI says it supports California Senate Bill 1119, which it describes as establishing age-appropriate safeguards for young people using AI. The post lists age determination, pre-release safety assessment, independent…
A milestone in expanding access to AI
OpenAI reports that ChatGPT Ads reached a US$1 billion annualized revenue run rate in less than 200 days after launch. It says the platform is used by tens of thousands of advertisers, is available in more than 40…
Our decision on Cursor following its acquisition by SpaceX
OpenAI says it notified SpaceX that it intends to wind down its contract providing OpenAI models to Cursor, with a proposed shutoff date of November 12, 2026. The company attributes the decision to the change of…
Page 3 of 47
