Last week we released Claude Sonnet 4.5. As part of our alignment testing, we used a new tool to run automated audits for behaviors like sycophancy and deception. Now we’re open-sourcing the tool to run those audits.
RESEARCH
-
Petri: Automated Agent Tool Audits AI Models Across Scenarios
By
–
It’s called Petri: Parallel Exploration Tool for Risky Interactions. It uses automated agents to audit models across diverse scenarios. Describe a scenario, and Petri handles the environment simulation, conversations, and analyses in minutes. Read more:
-

Bytedance GRPO Paper Challenges Optimization Approach Efficiency
By
–
Following an intense weekend of GRPO discussion, Bytedance put out a paper on why GRPO is NOT optimal! Similar to the classic knapsack problem, they suggest that each exploration has a “value” and “cost”, which needs to be adaptively distributed. This yields 20-40% more signal
-
Compile Effects on Neural Network Layers and MoE Architecture
By
–
Interesting, interesting. I thought compile would rather affect the nn.Linear layers vs nn.Embedding as the latter is a lookup and not doing any actual math/compute. Curious now and have to run some experiments later. But since you mentioned MoE, also I am confused now where you
-
CiteCheck Benchmark Standards for Precise Claim Verification
By
–
The goal of a CiteCheck benchmark should not be to check, "Does the cited document say something *related* to the claim?", but "Does the document state or very directly support the *exact* claim it's being cited about?" Here are two failures from the last generation of LLMs that
-
LLMs struggle with verifying cited sources and references
By
–
I've previously found LLMs to suck at "Track down cited pages/references and see if they support the citer's claim." LLMs hallucinate what the cited document says, if the citer's claim sounds LLM-plausible. I wish a CiteCheck benchmark for this kind of task would get put
-
G7 Semiconductor Group Addresses Supply Chain Resilience
By
–
G7 Semiconductor Point of Contact group meets in Germany on R&I and supply chain resilience | Shaping Europe’s digital future https://
digital-strategy.ec.europa.eu/en/news/g7-sem
iconductor-point-contact-group-meets-germany-ri-and-supply-chain-resilience
…
#digitaleu #Semiconductors #SupplyChain #innovation #technology @DigitalEU @elaniazito @ArturHabant @BetaMoroney @AkwyZ @enilev -

Hugging Face Transformers: Managing 1M Lines Python Code
By
–
With 1,000,000+ lines of Python and 400+ models How do we keep @huggingface Transformers sane? > Our design tenets: one-model-one-file, modular shards, minimal API, no breakage Peek under the hood on one of the biggest and most-used open ML codebases out there
-
AI Systems Solving Voice Recognition and Cocktail Party Problem
By
–
maybe that’s the use case for current/next gen AI systems. I mean voice inputs have been pretty much solved years ago. But there are still false positives, and newer systems may solve the cocktail party problem here.
-
Introducing Continuous Thought Machines New Framework
By
–
Introducing Continuous Thought Machineshttps://t.co/Fjwl82PZ9E
— hardmaru (@hardmaru) 6 octobre 2025
(Blog post from earlier this year.)
Summary tweet:https://t.co/bOGIDVDZodIntroducing Continuous Thought Machines https://
sakana.ai/ctm/
(Blog post from earlier this year.) Summary tweet:
