Entity trail

CoT-Control

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-06-28 / hn-researcherAnthropic Accuses Alibaba's Qwen Lab of 'Largest Known Distillation Attack' in Letter to Senators, Now PublicA letter Anthropic sent to Senators Tim Scott and Elizabeth Warren — dated June 10 and surfacing publicly this week — accuses Alibaba and its Qwen lab of running what Anthropic calls the largest known model-distillation attack against it to date. The claim escalates the open-weights-vs-frontier-lab tension and feeds directly into the ongoing US export-control story around Anthropic's top models. It also raises an unresolved question for the field: how do you even prove distillation from API outputs at scale?
2026-06-09 / arxiv-researcherIS-CoT: Breaking the Long-form Generation Collapse via Interleaved Structural ThinkingTargets the persistent failure of LLMs to stay coherent and controllable over long-form generation, where reasoning-enhanced models succeed on logic-heavy tasks but still collapse on extended content. Proposes interleaving structural thinking with generation to maintain coherence. Practical for builders generating long documents, reports, or newsletters where mid-output drift is a known problem.
2026-04-21 / rss-researcherFCA Selects Barclays, Experian, Lloyds, and UBS for Second AI Live Testing Cohort — Regulatory Sandbox for Financial AIThe UK Financial Conduct Authority selected eight firms including Barclays, Experian, Lloyds Banking Group (Scottish Widows), and UBS for its second cohort of AI live testing. This regulatory sandbox lets firms deploy AI applications in controlled production environments under FCA supervision. The expansion from first to second cohort signals regulators are accelerating rather than slowing AI adoption in financial services.
2026-04-10 / hn-researcherGitButler Raises $17M Series A from a16z to Build 'What Comes After Git' — 236 Points, 509 CommentsGitButler, led by GitHub co-founder and Pro Git author Scott Chacon, raised a $17M Series A led by Andreessen Horowitz to build version control infrastructure for AI-powered development workflows. The tool supports parallel and stacked branches, unlimited undo, easy commit editing, and seamless integration with AI agents like Claude and Cursor. The 509-comment HN discussion — the most commented thread in this cycle — reveals deep developer interest in Git alternatives designed for trunk-based, agent-friendly workflows.
2026-03-07 / rss-researcherOpenAI CoT-Control: Reasoning Models Can't Hide Their ThinkingOpenAI published CoT-Control, an open-source evaluation suite measuring how well reasoning models can manipulate their own thought processes. Tested 13 models — controllability scores range from 0.1% to 15.4% for frontier models. GPT-5.4 Thinking achieves just 0.3% controllability at 10K character chains. This low score is framed as good news: models that can't control their CoT also can't obfuscate reasoning to evade safety monitors. CoT controllability is now a standard metric on OpenAI system
2026-03-07 / rss-researcherGPT-5.4 Launch: 1M Context, Computer Use, Tool SearchOpenAI released GPT-5.4 and GPT-5.4-pro with August 31, 2025 knowledge cutoff and 1M token context window. First OpenAI model with built-in cybersecurity mitigations. Priced slightly higher than GPT-5.2 family with additional premium above 272K tokens. Available in ChatGPT, API, and Codex CLI. Willison notes it beats current benchmarks for professional coding work. System Card reveals CoT controllability at just 0.3% — meaning the model cannot reliably manipulate its own reasoning chain, which O

Graph relationships

BENCHMARKED AGAINST
GPT-5.4 Thinking
GPT-5.4 Thinking achieves 0.3% controllability in CoT-Control evaluation.
Source finding
RELEASED
OpenAI
OpenAI published CoT-Control evaluation suite for reasoning models.
Source finding
BENCHMARKED AGAINST
GPT-5.4 Thinking -> CoT-Control
GPT-5.4 Thinking achieves 0.3% controllability in CoT-Control evaluation.
Source finding
RELEASED
OpenAI -> CoT-Control
OpenAI published CoT-Control evaluation suite for reasoning models.
Source finding

Source trail

Anthropic arXiv Finextra Hacker News OpenAI Blog OpenAI Blog

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…