Notes on AI, research, and the work.
234 pieces on generative and agentic systems, the research underneath, and what it takes to ship them.
- Oct 2026 · 5 min
Learn2Play Bench: do LLM agents actually learn from experience?
A new benchmark of text games with novel, counterintuitive rules separates real learning from recalled pretraining. The practical results for agent builders: raw experience logs are competitive with, and often better than, distilled rules, and the harness moves scores as much as the model does.
llm-agentsagent-evaluationagent-memorybenchmarksRead - Oct 2026 · 5 min
MiMo-V2.6 and the three axes of scaling agentic RL
Xiaomi's MiMo-V2.6 report treats RL as something you scale along three separate axes: training compute, environment diversity, and grader compute. The most useful parts for practitioners are the reward-hacking defenses and the groupwise grading that ranks passing solutions instead of rewarding them all equally.
reinforcement-learningagentic-aireward-hackingmixture-of-expertsRead - Oct 2026 · 6 min
DecepEval: measuring when agents lie, not just whether they can
DecepEval runs 1,532 agent tasks twice, once neutral and once with a single added pressure, incentive, opportunity, or conflict, and measures how much deception rises. The paired delta is the useful part for anyone building agents, and the headline result is that a low baseline tells you little about behavior under inducement.
llm-agentsai-safetydeceptionevaluationRead - Oct 2026 · 3 min
Agent Lightning v1.0: RL training on the agent you actually ship
Microsoft Research Asia rebuilt Agent Lightning around 'Harnessed Agentic RL': your deployed agent harness stays untouched, and a thin LLM proxy turns its model calls into training data. It is a small, practical framework release rather than a frontier result, and that is why it is worth a look.
agentic-aireinforcement-learningai-agentsmlopsRead - Oct 2026 · 9 min
It improves itself. It doesn't grade itself.
What recursive self-improvement is in plain terms, how to keep a human on the loop instead of in it, and the lessons that cost me the most: separate the ruler from the thing it measures, measure first, fail loudly, and earn autonomy on evidence.
rsiself-improvementhuman-on-the-loopevaluationRead - Oct 2026 · 5 min
EmbeddingGemma 2 brings text, image, audio, and video into one small embedding model
Google DeepMind's EmbeddingGemma 2 extends last year's text-only embedder to code, images, video, and audio in a 740M-parameter, Apache 2.0 model built for on-device retrieval. It's an incremental family release, but a useful one for anyone building local or open-weights RAG.
embeddingsmultimodalon-device-airagRead - Oct 2026 · 7 min
UNREAL: one frozen LLM as both retriever and long-context reader
NVIDIA's UNREAL trains under 500K parameters on top of a frozen decoder-only LLM so the model ranks evidence itself, from a 21M-chunk Wikipedia index down to chunks inside a single long prompt. The multi-hop recall gains are large; the real question is how much survives outside Wikipedia QA.
retrieval-augmented-generationlong-contextllm-researchinformation-retrievalRead - Oct 2026 · 6 min
MedGemma and the case for a small, open medical base model
Google's MedGemma paper in Nature Medicine argues that a medically tuned 4B/27B open model, plus a medical vision encoder, is a better starting point for clinical AI than either a general base model or a closed frontier API. Here is what the evidence supports, and what it does not.
medical-aivision-language-modelsopen-modelsfoundation-modelsRead - Oct 2026 · 3 min
Quine: Microsoft Research's bet on a world model of biology
Microsoft Research introduced Quine, a multimodal world model of biology inside a harness that connects tools, literature, scientists and the wet lab. The goal is well specified and the pancreatic cancer case study is encouraging, but the announcement reports one result without numbers, not measured performance.
ai-for-scienceworld-modelscomputational-biologydrug-discoveryRead - Oct 2026 · 4 min
MotorMind and the case for scaffolding general VLMs to run robots
A new paper wraps an off-the-shelf VLM in mid-level actions, deterministic control and async feedback, and reports far better zero-shot manipulation than earlier baselines. The pattern is familiar from agentic software; the evidence so far is narrow.
roboticsvision-language-modelsagentic-aizero-shotRead - Oct 2026 · 4 min
Offloading robot inference: a systems finding worth taking seriously
Microsoft Research measured what happens when a robot's AI inference runs on edge or cloud GPUs instead of onboard hardware. Task success and battery life improved, and the idea is now packaged in the Physical AI Toolchain.
physical-airoboticsedge-inferencekubernetesRead - Oct 2026 · 6 min
Where-OPD: teaching multimodal models where to look with synthetic scenes
A paper applies on-policy self-distillation to multimodal models, giving the teacher textual spatial hints from procedurally generated scenes. It needs no annotation or external teacher and reports a 3.23-point average gain on real-world perception benchmarks.
multimodal-llmself-distillationsynthetic-datacomputer-visionRead - Oct 2026 · 4 min
OpenAI's distillation disclosure: when the model becomes the decryption oracle
OpenAI says it disrupted a coordinated adversarial-distillation campaign that got its models to decrypt their own protected reasoning when it was replayed from another conversation. What happened, what OpenAI changed, and what it means for anyone shipping reasoning models.
ai-securitymodel-distillationreasoning-modelsopenaiRead - Oct 2026 · 7 min
Private AI Compute gets a memory, and the privacy guarantee changes shape
Google is adding persistent, per-user memory to Private AI Compute, encrypted under user-bound keys and decrypted only inside attested enclaves. The design is careful, and its most important detail is the guarantee it openly gives up.
private-aiconfidential-computingagent-memoryai-securityRead - Oct 2026 · 7 min
Teaching audio-visual models to look and listen on demand
OmniSeek turns an omni LLM into an agent that fetches audio or video evidence mid-reasoning instead of ingesting the whole stream upfront, trained with synthetic trajectories and a reward that punishes single-modality shortcuts.
agentic-aimultimodalreinforcement-learningtool-useRead - Oct 2026 · 7 min
Gemini 4 Argon: a frontier model you mostly can't touch yet
DeepMind announced Gemini 4 Argon with a stack of benchmark wins and a 1M-token output limit, but it's shipping to a narrow set of trusted cyber defenders, not developers. Here's what the announcement actually says, and what it conspicuously leaves out.
gemini-4-argongoogle-deepmindfrontier-modelsai-benchmarksRead - Oct 2026 · 6 min
Your eval pipeline has an undocumented input: today's date
A new paper shows that the current date silently injected into LLM system prompts shifts benchmark scores by up to 14% on math reasoning and reshuffles leaderboard rankings — a reproducibility hole most eval harnesses never control for.
llm-evaluationreproducibilitybenchmarkingsystem-promptsRead - Oct 2026 · 7 min
DeepMind watermarks AI-designed proteins with SynthID Bio
DeepMind has extended its SynthID watermarking approach from images and text to proteins, embedding a detectable signal in both the amino acid sequence and the AlphaFold 3-predicted structure without disrupting biological function. It's a research proof-of-concept, not a deployed screening tool, but it closes a real gap in biosecurity.
ai-biosecuritysynthidprotein-designdeepmindRead - Sep 2026 · 7 min
The bottleneck in long-horizon agents isn't the model, it's the control loop
A new paper from Meta and collaborators shows that adding a lightweight controller that reasons over a compact run summary, rather than the full transcript, beats direct-control agents and even production coding agents like Codex and Claude Code on long-horizon benchmarks.
agentic-aillm-agentsmeta-reasoningai-researchRead - Sep 2026 · 6 min
Entities as anchors: giving agentic search a map of the corpus
A new paper, Follow the Entities: A Corpus Map for Agentic Search, builds an offline graph of recurring entities that links documents together so a search agent stops rediscovering the same relationships on every query — and uses fewer tokens doing it.
agentic-searchretrieval-augmented-generationentity-resolutionllm-agentsRead - Sep 2026 · 7 min
Curate memory when the task is known, not when the task ends
A new paper argues that agent memory systems have been curating at the wrong moment — distilling experience into fixed summaries right after a task ends, instead of waiting until a new task arrives and shaping memory around it. The fix trains cleanly on immediate task success instead of a delayed reward signal.
agent-memoryllm-agentsreinforcement-learningresearchRead - Sep 2026 · 6 min
MindTopo shows vision-language models can see structure but can't hold onto it
Microsoft Research's new MindTopo benchmark finds that multimodal models recognize topological relationships like connectivity and enclosure in a static image, but lose track of them once they have to act on a scene over multiple steps.
vision-language-modelsspatial-reasoningbenchmarksroboticsRead - Sep 2026 · 6 min
Gemini 3.8 Live gets a face, not a new brain
Google DeepMind's Live Avatar adds a lip-synced, expressive video layer on top of the existing Gemini 3.8 Live dialogue model — a front-end feature for enterprise agents, not a new capability at the model level.
geminigoogle-deepmindvoice-aimultimodal-aiRead - Sep 2026 · 6 min
OpenAI tightens prompt caching for GPT-6, and it matters more than it sounds
OpenAI's GPT-6 prompt caching update raises hit rates, adds explicit cache breakpoints and a diagnostics dashboard, and pushes cached-token discounts to as much as 90% — a plumbing change that quietly reshapes how you should design long-running agents.
openaigpt-6prompt-cachingllm-infrastructureRead - Sep 2026 · 10 min
Jev picks. It doesn't write.
A visual walk through TypeSafe's Jev: what it is, what's actually known about how it's built, what "zero hallucination" really means, and where I plan to use it: my blog scout and my self-improving harness.
jevsystem-one-modelscalibrationevaluationRead - Sep 2026 · 7 min
Recursive self-improvement of agent harnesses needs a regularizer
A new paper, RRSI, shows that letting an agent rewrite its own harness overfits to the training benchmark unless you constrain how it proposes and selects edits — the fix combines an annealed edit budget, novelty-seeking search, and a critic-plus-pruner selector.
agent-harnessself-improvementregularizationllm-agentsRead - Sep 2026 · 6 min
IntBMoE: what if a mixture-of-experts model didn't have to pick two out of three
IntBMoE decouples participation, execution, and materialization in Mixture-of-Experts by composing experts with a hypernetwork before routing sparsely to a fixed codebook of blocks — and it's already running in AMap's recommender at scale.
mixture-of-expertsmodel-architecturerecommendation-systemsproduction-mlRead - Sep 2026 · 7 min
Register tokens: giving diffusion language models a bounded memory for reasoning
A new paper trains diffusion language models to carry reasoning across generation chunks in a fixed-size set of register tokens instead of the growing text context, beating a discrete-text-carry baseline by up to 19.5 points on code and 8.5 on math.
diffusion-language-modelsllm-researchreasoningcontext-managementRead - Sep 2026 · 7 min
Verifiable ground truth for social reasoning in LLM assistants
Google Research's Fuse framework gives social-reasoning evals an actual verifiable answer by hiding a motive inside a multi-agent simulation, then uses it to show assistants are swayed by biased framing and don't reliably get better with longer conversations.
llm-evaluationsocial-reasoningmulti-agent-simulationgoogle-researchRead - Sep 2026 · 6 min
SoL-Pi: an automated search over agent-harness design cuts token cost by a third — with real caveats
A new NVIDIA/NTU/MIT paper runs an automated search over agent-harness design choices rather than the model itself, surfacing four mechanisms that cut coding-agent token traffic up to 49% and cost by about a third — at a real, if small, accuracy cost the abstract doesn't emphasize.
agent-harnesstoken-efficiencyautomated-researchllm-agentsRead