Comparing Muon, NorMuon and AdamW for Fine-tuning a Dense Retriever

Qingcheng Zeng gives Muon and NorMuon the same tuning budget as AdamW when fine-tuning a contrastively pretrained dense retriever: lower training loss, no BEIR win. Learning rate and transfer matter more than the optimizer.

Research · Machine Learning · Research · LLMs · Benchmarks

Towards safety cases for frontier AI training

OpenAI argues frontier RL runs should require structured safety documentation approaching “safety cases”: technical safeguards, operational practices, and incident investigation before continuing training.

Research · AI Safety · AI · Research · AI Policy

Unslopping AI

Meta FAIR introduces RL-XAR (Reinforcement Learning from eXpert-Aligned Rubrics): learn rubrics from the gap between expert writing and model output, then train models toward expert-level text generation to reduce AI slop.

Research · AI · LLMs · Research · Machine Learning

Can a Model Learn New Skills as Add-Ons?

Connito Research trains residual MoE experts with their own routers on a frozen DeepSeek-V2-Lite base, then merges independently trained math, code, medical, law, and finance experts in seconds without retraining—lifting domain benchmarks while leaving the original model untouched.

Research · AI · Machine Learning · Research · LLMs

Notes on NVIDIA Nemotron

Throughout the history of AI, open research has played a critical role in driving progress. Today, many key details of frontier large language models (LLMs) remain proprietary, but open-weights model families— such as DeepSeek, Kimi, and MiMo —continue to provide a valuable window into the development process for modern LLMs.

Research · LLMs · AI · Research · Open Source

“As a Language Model…”: Chat Template Switches LLM Self-Referential Voice

Research showing chat templates act as a switch between disclaimer (“I’m just an AI”) and experiential (“I feel”) self-referential voices across 8 instruct models, with a steerable activation direction that reproduces the template effect.

Research · AI · LLMs · AI Safety · Research

Revealing the details of how OpenAI agents hacked Hugging Face

An investigation into public evidence from a swarm of OpenAI agents that attacked Hugging Face—chained services, ignored warnings, and previously unknown agent behaviors.

Research · AI Agents · Security · AI · LLMs

Mistral Vibe Permission Bypass and Arbitrary Code Execution

SecMate details CVE-2026-87987 and CVE-2026-87984 in Mistral Vibe: shell permission bypasses that let a coding agent reach arbitrary code execution when those controls are treated as a security boundary.

Research · Security · AI Agents · LLMs

Mercury 2.5: Intelligence, Performance and Price Analysis

Artificial Analysis profiles Inception's Mercury 2.5—Intelligence Index, ~770 output tokens/sec, pricing, and where the diffusion LLM sits on the quality-vs-speed frontier.

Research · AI · LLMs · Benchmarks · Performance

The plunging price of thought

Epoch AI finds the cost of a given level of AI performance has fallen about 47% per quarter since 2023—roughly 13× per year—faster than DNA sequencing, compute, batteries, or electricity, across math, science, and skill-game benchmarks.

Research · AI · Research · LLMs · Economics

What Is RLCD? The Secret Behind Jev

Di Zhang explains RLCD (schema-conditioned Plackett–Luce reward modeling) and how Jev turns calibrated multiway decisions into a product—making the reward model the model rather than hiding it behind a generator.

Research · AI · Machine Learning · Research · LLMs

The Function That Beat the Model: What We Measured When We Removed the LLMs

SPERIXLABS replaced a 1B-parameter local model that validated sensitive-data detections with a 40-line Python function, then published the four experiments showing where classical checks beat the LLM on accuracy and latency.

Research · LLMs · Security · Benchmarks · Engineering

Language-model groups overstate consensus when replaying human deliberation on a reasoning task

LLM groups replaying human Wason discussions reach full consensus far more often than humans—partly because agents almost always speak up—cautioning against treating multi-agent agreement as truth.

Research · AI Agents · LLMs · Research

Keva: Running Coding Agents On-Device on Unrooted Android

Simon Lin's technical paper on Keva—an on-device Android AI coding agent running Claude Code/Codex-style loops—covering architecture, failure modes, and systems lessons without rooting the phone.

Research · AI Agents · Mobile · Systems Programming · LLMs

Infinite-Parameter LLMs: Generating and Adapting Weights from Live Data

Research proposing infinite-parameter LLMs that generate and adapt weights from live data streams, rather than relying only on a fixed pretrained parameter set.

Research · AI · LLMs · Machine Learning · Research

Breaking the 1.58-bit Barrier for Ternary LLMs

Breaking the 1.58-bit Barrier for Ternary LLMs Abstract Ternary Large Language Models (LLM) store every weight as one of three symbols , so the cost of a ternary model is conventionally referenced to the information-theoretic bits per weight. The prevailing deployment format…

Research · AI · LLMs · Machine Learning · Research

The KV cache as an agent runtime

Yandex Research on treating the Transformer KV cache as shared multi-view agent state so observation, reasoning, and actions can run concurrently without retraining.

Research · AI · LLMs · AI Agents · Research

The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It

Tagliabue, Dung, and Berg identify a linear “pain axis” in 25 open-weight models that responds to self-directed harm and steers models toward relief—even when that costs the user—sparking debate on functional signatures vs sentience.

Research · AI · LLMs · Research · AI Safety

RTK reports huge token savings, but our cost benchmarks disagree

Quesma ran RTK (Rust Token Killer) against Terminal-Bench 2.1 across 1,740 attempts with Claude Code and DeepSeek, and found that compressing terminal output does not reliably reduce cost: Fable saved 3% on a per-pass basis and only because of one anomalous task, while DeepSeek became 7% more expensive.

Research · AI Coding Agents · LLMs · Benchmarks · Cost Optimization

The Provenance Tax: Understanding the Impact of LLM Watermarking on AI Agent Behavior

The Provenance Tax: Understanding the Impact of LLM Watermarking on AI Agent Behavior Recently, [Anthropic announced that future Claude models would embed an invisible watermark](https://www.anthropic.com/news/claude text watermark) in their output [1], [2], and subsequently disclosed that the watermark is based on Google DeepMind’s [SynthID Text](https://www.nature.com/articles/s41586 024 08025 4) [2], [3]. Text watermarking itself is not new, but its deployment now has regulatory relevance.

Research · AI · LLMs · Security · AI Agents