counterfactual paths cut credit assignment noise in llm reasoning
a new method uses counterfactual trajectory comparisons to turn sparse terminal rewards into step-level signals, stabilizing reinforcement learning for multi-step llm reasoning.
author
baris
Baris operates SummarizedData and maintains its sources, publishing rules, and automation. Read the editorial process.
adversarial action removal breaks self-play rl agents
attackers who delete legal actions before an agent decides cause severe and lasting damage across multiple games and algorithms.
signmuon: 1-bit matrix-aware optimizer for distributed training
signmuon combines 1-bit sign communication with muon's matrix-aware updates to reduce distributed training bottlenecks.
skim speeds up web agents with speculative execution
skim uses site-specific templates to skip heavy ai steps for most web tasks, falling back to full agents only when needed.
agentwall adds runtime safety for local ai agents
agentwall intercepts agent actions before execution, enforcing declarative safety policies to prevent harmful operations on local machines.
stopping local ai agents early saves energy
agentstop reduces energy use in local ai agents by predicting task failure early, cutting token waste and battery drain on consumer devices.
risk-aware policy learning for offline bandits
a new method learns safe decision rules from logged data using general risk measures like cvar, with strong theoretical guarantees.
six months of llm progress in five minutes
a lightning talk summary of major llm developments from november 2025 to may 2026, including coding agent breakthroughs and open-weight model advances.
anthropic buys stainless, cutting off rival ai labs from key dev tools
anthropic acquires stainless, the sdk automation startup used by openai, google, and cloudflare, and will shut down its hosted products.
sandboxaq puts drug discovery ai into claude
sandboxaq integrates its physics-grounded quantitative models into anthropic's claude, letting researchers run complex simulations through plain language prompts.
daily brief: 2026-05-18
musk's openai lawsuit dismissed, new tools for local ai and ocr, and research on hidden bias in language models.
paddleocr 3.5 adds transformers backend
paddleocr 3.5 lets developers run ocr and document parsing models with a hugging face transformers backend, reducing integration friction for rag and document ai workflows.
musk loses openai lawsuit on statute of limitations
a california jury found elon musk's claims against openai and sam altman were filed too late, removing a major legal threat before openai's reported ipo.
the hidden skill gap beyond sql and python
data jobs now demand data modeling, performance optimization, infrastructure awareness, and practical ai skills beyond basic sql and python.
pytorch 2.11 fixes aarch64 gpu wheel installs
pytorch 2.11 now publishes cuda-enabled wheels for aarch64 linux on pypi, removing the need for custom indexes and workarounds when deploying on nvidia grace hopper and grace blackwell systems.
pytorch executorch mlx delegate for apple silicon gpus
a new executorch backend enables gpu-accelerated inference on apple silicon macs using apple's mlx framework, with broad model and quantization support.
latent bias in llms can flip decisions
instruction-tuned language models show fair outputs but retain biased internal representations that can reverse decisions when activated.
fine-tune cosmos predict 2.5 with lora for robot videos
a guide to parameter-efficient fine-tuning of nvidia cosmos predict 2.5 using lora and dora for generating synthetic robot manipulation videos.