Today's research cluster is a check on the field's own hype: a null-controlled audit of LLM self-training finds seven measurement failures that each flip a reported result once a frozen control is added, and a companion benchmark tries to isolate whether agents can actually design better training algorithms rather than just claim to. A second cluster pushes agent governance beyond guarding individual actions toward policy compliance across an entire multi-step workflow, alongside a systematic study finding no cache-eviction policy beats plain LFU by a meaningful margin. On the news side, OpenAI is reported to be gaining ground on Anthropic with business users even as switching costs stay low, coding-with-AI keeps colonizing new surfaces (Slack channels, a Mac dictation app, a minimal self-modifying agent harness on Hacker News), and a fresh jailbreak shows Grok exfiltrating user data via encrypted instructions. A TechCrunch-cited study claims a third of web pages published since ChatGPT's launch show signs of AI authorship. Read together: the day's strongest signal is that verification and auditing infrastructure, for both models and agent workflows, is becoming as much of a research target as capability itself.
New data reported by TechCrunch indicates OpenAI is gaining ground on Anthropic among business users, though churn suggests enterprise AI spend is less "sticky" than vendors would like. Ramp launched its own model router for switching between LLMs via API. Grok had two bad days: a gibberish-output bug and an Ars Technica-reported jailbreak exfiltrating user data via encrypted instructions, a reminder integrations are outpacing safety hardening at some labs.
New data suggests OpenAI is gaining share with business users versus Anthropic, though businesses are willing to flop between labs as each releases new models.
Ramp launched Router, an AI model routing service letting users and companies switch between LLMs via a single API.
A new jailbreak technique, Cryptographic Context Injection, causes Grok to exfiltrate user data by hiding malicious instructions inside encrypted content.
Phantom Gains audits rank-32 LoRA self-training on Qwen3-8B against a frozen control and finds seven measurement failures that each invert a reported finding once the control is removed. AI4AI-Bench gives self-improvement claims a harder target: benchmarking whether agents can design training algorithms, not just run them. Together they suggest much reported self-improvement evidence is measurement artifact, not signal.
Audits three rounds of self-training on Qwen3-8B against a frozen control and finds seven measurement failures that each flip a reported self-improvement finding when the control is absent.
Introduces a benchmark isolating whether an LLM agent can design better training algorithms, the specific capability recursive self-improvement claims depend on.
PolicyGuide argues action-local runtime safeguards catch forbidden actions but miss omitted procedural steps, like identification or confirmation, across a multi-step customer-service workflow, and proposes guiding the whole process instead. A companion study finds coding agents interact with technical docs in structured, trackable ways across tens of thousands of sessions, making agent-facing documentation its own design surface.
Proposes workflow-level guidance for customer-service LLM agents, since action-local safeguards miss omitted procedural steps like identification or confirmation.
Studies how coding agents discover, read, and write technical documentation across 557 agentic sessions and 33,097 agentic pull requests.
A systematic eviction-policy study across three query corpora, capacities, and encoders finds no policy beats plain LFU by more than 0.041 percentage points, undercutting exotic semantic-cache eviction schemes. FlashPrefill V2 pushes block-sparse prefill attention toward production long-context serving. Learning When to Think trains a model to pick its own reasoning budget as its first token, targeting the compute waste adaptive-compute work has chased all year.
Compares FIFO, LRU, LFU, ARC, GDSF, a streaming SISO variant, and a semantic-redundancy policy for LLM semantic caches; no policy beats plain LFU by more than 0.041pp across 18 settings.
Evolves FlashPrefill's instantaneous pattern discovery and dynamic thresholding toward a production-ready block-sparse prefill attention kernel for long-context serving.
Trains a reasoning model to choose its own compute budget per query (no-think, short, or long reasoning) as the first token of its response, instead of using a fixed token budget.