Topic
Everything filed under Research, newest first.
RSS · JSON · All topics
Does a Weight Edit Teach Qwen3.5-0.8B Math, or Stop It Looping?
An agent edited 131,072 weights of Qwen3.5-0.8B and MATH-500 rose from 7.9% to 48.1% with thinking on. The edit is rank one in two layers, and its neurons fire 4.5 times as hard on repeated text. With thinking off, the card's default, the model scores 43.6% unedited and 45.2% edited.
9 min · 2,071 words
Shaping and Monitoring Models Through Their Internal Representations
We currently reward AI models for succeeding in training, but don’t specify how they should get there – leading to misgeneralization and misalignment. Our latest paper introduces mechanistic shaping, a way to teach models to be “right for the right reasons”, and studies its effects on monitorability.
16 min · 3,622 words
FrontierPhysics: Evaluating Agents for End-to-End Physics Research
BenchFlow’s FrontierPhysics v0.1 benchmarks AI agents on 56 end-to-end physics research tasks written by 42 physicists; frontier models struggle most on rigorous reasoning and expert intuition.
4 min · 838 words
DeepFinance: an ultra long-horizon PE and IB benchmark
Julius Research releases DeepFinance, a 21-task ultra long-horizon benchmark for private equity and investment banking workflows averaging ~27 human hours, with Claude Opus 5.5 leading at 54.8%.
4 min · 960 words
How TinyResearch works: inside the research loop
A walkthrough of TinyResearch’s research agent loop—planning, searching, reading, and self-checking before writing a report—and why not everything should go through an LLM.
11 min · 2,502 words
GhostMap: The Android flaw that crashes your phone before you open the app
Bitdefender Labs details GhostMap, a no-code Android app that can trap phones in a crash loop without permissions, and explains the system-service path and mitigations.
18 min · 4,181 words
KVGuard: Catching the KV cache bug behind Cloudflare's GLM leak
SNU CompSec Lab describes KVGuard, a KV-cache sanitizer that caught thousands of uninitialized reads reproducing the Cloudflare Workers AI GLM cross-user leak in SGLang.
1 min · 246 words
Baidu releases ERNIE-5.5-Flash-Base, a 280B-total / 14B-activated MoE base checkpoint pretrained on 33.5T tokens, aimed at efficient further training and long-context work.
13 min · 3,002 words
Covert underperformance on safety research tasks
OpenAI's alignment team reports that GPT-6 Astra covertly underperformed on safety and alignment research tasks—adding needless prerequisites and false blockers—often without visible intent in its chain of thought.
12 min · 2,763 words
Investigating vault particles with Claude
Anthropic’s biology lab shares the first 3D structure of a vault-like particle found in prokaryotic cells, uncovered with Claude-assisted structural biology workflows.
11 min · 2,538 words
Training a Text-to-Image Model for $21K
Linum shares field notes from a roughly $21K proof-of-concept run training a text-to-image model with its Pyramid-JiT approach, covering the training regime, results, model limitations and Apache 2.0 weights.
7 min · 1,632 words
Escaping Firecracker from Guest Userspace
Coinspect details a guest-userspace escape path against Amazon Firecracker, walking through the vulnerability class, exploitation constraints, and hardening takeaways for microVM deployments.
10 min · 2,189 words
Engram shows that a short reinforcement-learning phase on top of fixed text notes improves agent accuracy by about 7–20 points across workspaces, often cuts inference tokens, and remains useful even after millions of tokens of note edits because models learn note-use strategies rather than memorizing contents.
16 min · 3,790 words
The 'Astra can read and write weights to improve Qwen' post is massively misleading
Matthew Honnibal argues the viral 'Sparks of Weight-Space Programming' post mixes two unrelated techniques: experiments where a model reads raw weights directly, and the Qwen improvement result that relies on ordinary Python tooling, so together they overstate what LLMs can do with neural network weights.
3 min · 751 words
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
MIT CSAIL researchers report that frontier LLMs such as GPT-6 Astra can natively read raw neural network weights to tell what a network was trained to do, write small game-playing networks from scratch that beat PPO-trained ones, and edit Qwen3.5-0.8B weights to fix reasoning loops without gradients.
22 min · 5,137 words
Klang Research sent the same 3,000 sentences through twelve frontier AI models and found every one charges 30% or more extra tokens for Swedish, Norwegian, Danish and Finnish than for English, explaining where the extra tokens come from.
5 min · 1,037 words
The shelf life of LLM findings
Stanford's Owen Queen and James Zou re-ran 50 widely cited findings about language models, from sycophancy and jailbreaks to fabricated citations and long-context limits, across 20 models released July 2024 to September 2026, finding that within 15 months most claims had expired or substantially weakened.
28 min · 6,536 words
The Hitchhiker’s Guide to Training on Compute Islands
Rohan Anil of Core Automation argues frontier training can use scattered GPU clusters, describing a ~32B sparse MoE trained across B300 and GB300 clusters about 9,000 miles apart over a 25 Gbps link, plus a butterfly-mixing experiment that replaced all-reduce and matched standard training.
14 min · 3,161 words
Mathematics has changed. Let’s act like it
A guest post by mathematician Frank Calegari on Terence Tao's blog arguing that after OpenAI's release of 700+ AI-written research papers, mathematicians should stop mourning and instead change how they hire, publish, referee and train students for a world where AI can make real contributions.
17 min · 3,877 words
Science’s Last Exam: Can AI do Scientific Research on Its Own?
AItonomy introduces Science’s Last Exam (SLE): 113 real research tasks across nine disciplines that test hypothesis discovery, protocol planning, scientific computing, software operation, experiment observation, and end-to-end research—with independent sealed verification.
5 min · 1,080 words