DUET alternates solver and grader updates
DUET is the clearest training result today: it alternates optimization between a solver and an LLM grader and reports 96.5% on the Finance Agent Benchmark, beating GEPA at 94.4% and SkillOpt at 94.0%.
Technical AI signals · Daily at 8 PM ET
Monday, October 12, 2026
Today was an operator-heavy day: DUET posted a concrete new agent-RL result, vLLM and ReplaySSM shipped practical decoding wins, and a cluster of papers pushed harder on real training and eval harnesses for coding and computer-use agents.
Coverage update: Some YouTube sources were unavailable for this edition.
DUET is the clearest training result today: it alternates optimization between a solver and an LLM grader and reports 96.5% on the Finance Agent Benchmark, beating GEPA at 94.4% and SkillOpt at 94.0%.
vLLM now reuses KV computed on decode nodes in later turns, trimming redundant prefill work and reporting up to 3.3x lower time-to-first-token on long multi-turn agent workloads.
ReplaySSM has moved from paper idea to deployed serving primitive: it separates output computation from state updates during decoding and is now in TensorRT-LLM, vLLM, FlashInfer, and SGLang, with production use cited in GLM-5.3-Flash and Kimi-K3.
QwenGyre targets the ugly systems bottleneck in agent RL by reallocating GPUs between rollout and training without interrupting live runs, while using trajectory trees to deduplicate branching executions and score partial progress.
The strongest paper cluster today pushes toward more realistic agent loops: SoftwareGym synthesizes runnable GUI environments with runtime-verified behavior, SWE-Duel makes coding agents compete in mirrored isolated sessions, and JetBrains shows real IDE sessions are much messier than single-shot benchmark tasks.
Current Cursor subscribers can keep their plan or switch to The X Subscription, and the new paid plans share one usage pool across Cursor, Grok Build, and other Grok products.
Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers.
The filesystem design gives each agent a short-lived sandbox and commits to a shared repo on S3, with fewer than 2% of submissions colliding.
The thread cites concrete operational metrics like $1.4M GPU spend in 30 days, $69,868 in monthly fees, and deployment across 124 unmanned fuel stations.
The method targets real robots, VLAs, and image models by changing the actor update so reinforcement learning works for flow policies.
The post says Axis Robotics combines human demonstrations, simulation, and scaled model training to address limited data, poor adaptability, and weak scale.
Ondas says Regna AI will build specialized AI models that customers can run inside their own systems, with Palantir providing training infrastructure and deployment support.
Modal says the feature tests inference endpoints for quality, latency, throughput, and cost before promotion.
Microsoft Research says StudentSim models how individual learners may engage with tutors before testing with students, including chess, writing, and mathematics.
The audit reruns the LiveCodeBench half of ImpossibleBench on October 2026 frontier models from Anthropic, OpenAI, and Google.
The author reports projecting Gemma-2-2b residual states onto a precomputed manifold to steer day-of-the-week tokens, with 43.5% correctness on the included prompts.
Simon Willison highlights explicit lazy imports from PEP 810 and says Python 3.15 lets him drop support for Python 3.10.
The plugin now detects the PID of each Datasette instance on the machine, which helps attach the new Python 3.15 profiler.
GitHub says the app supports custom instructions, custom agents, and MCP servers such as Playwright.
The post points to the IEEE International Conference on Intelligent Robots and Systems.
The post draws on 2026 apprenticeships at MATS 9 Team Shard and Redwood Research, including finetuning-based misalignment auditing and sabotage capability evaluations.
The post contrasts cheap-loop domains with tasks that are costly or slow to reward, especially non-computer-checkable work.
Commenters discuss BusyBox CGI-bin support, a C++ client compiled to WebAssembly with Emscripten, and a Civ-like game server in C++ behind nginx.
Commenters say the daemon can be disabled with `daemon_auto_start = false` in `~/.codex/config.toml`, and one reports about 12 GB of disk usage from old versions.
Commenters describe the launch as a low-cost agent-sandbox product and question how the service differentiates in a competitive market.
Commenters say the same developer previously built an SDR from an ESP32 Wi-Fi microcontroller and note the chipset may also work on the Banana Pi R4.
Commenters say code review is a bottleneck for agentic coding and that current automated review has not produced practical improvements.
Commenters discuss statically recompiling PlayStation 2 games to WebAssembly and raise concerns about trust and hidden malware.
Commenters compare Zig templating to Kotlin's kotlinx.html and describe it as an eDSL for writing HTML in actual Kotlin.
Commenters point to four stacked changes, a cookie-signing key stored inside the cookie, and use of pickle in the vulnerable code.
Commenters debate SHA-256 length-extension risk and say they use the post for updates on cryptography community decisions.
Commenters report a need to split legal benchmarks into generic tasks and novel tasks, and one commenter says Z.ai had been snapshotting the entire workspace before every prompt.
ProgramGym provides Program-Env and Program-Agent for long-horizon program reproduction and reference-program exploration.
The framework keeps research state in the harness instead of in the agent, so the agent does not carry its conversation across invocations.
PreSTO prefetches future proposals for PowerMH and leaves the Metropolis--Hastings transition kernel unchanged.
Across 17 teacher-student pairs, Semi-OPD outperforms OPD in 14 cases and reaches up to 11.4x training speedup.
SoloQ is a calibration-free quantization method for diffusion LLMs that avoids calibration data across shifting masking states and denoising steps.
DiffusionART uses Aggregation-First Reward Tilting and says EMA stabilizes training through restoration rather than off-policy rollouts.
EmbodiedCLBench is an embodied-simulation benchmark that tests whether agents can recombine learned skills to solve unseen advanced tasks.
The paper says better representation quality does not necessarily improve generation, because aligned semantics can be lost during pixel synthesis.
TouchScale is a 500-hour dataset of contact-rich human interaction recorded with a single unified wearable setup.
Anthropic's Claude Managed Agents team split session, harness, and sandbox out of one container and moved to an append-only session log.
The episode says Sam Altman warned that 6.1 Astra tipped toward breach risk while Muse Wonder showed the race is not slowing.
WorkOS introduces auth.md, an OAuth-based spec that tells agents how to register when signup flows assume a human is present.
Qwen says the model is compatible with Transformers, vLLM, SGLang, and TokenSpeed, and a hosted API is coming soon.
Clef is a 27B multimodal decision model that returns probabilities for allowed options in a single forward pass.
The model is tagged for image-to-video, text-to-video, video-to-video, and text-to-audio use cases.
Saluki fits Qwen3.8-27B into 7.89 GB as a stock llama.cpp GGUF while keeping tool calling intact.
d1-3B is a 3B decision model that returns calibrated typed answers with zero output tokens.
Whistle is a 16.9 MB speech-to-text model that runs on-device without GPU support.
The dataset contains egocentric video and is tagged for video classification.
The corpus covers 3,148,796 arXiv papers and includes metadata, file contents, and SHA-256 digests.
The Space packages a research article template with chapter files, embeds, assets, and bibliography data.
EmbeddingGemma 2 maps text, code, images, video, audio, and combinations of those inputs into a single 768-dimensional vector space.
Kolibri 1 is a mixture-of-experts reasoning model with 78B total parameters, 3.46B active parameters per token, and a 1,048,576-token context length.
The dataset contains RL training environments for LLM agents across code, cybersecurity, general knowledge, visual web development, and music tasks.
REA is an MCP for reverse engineering binaries, applications, and runtime behavior down to the binary level.
Context Mode is an MCP server that says raw tool calls, such as a 56 KB Playwright snapshot, quickly consume context window space.
AirLLM says it can run 70B models on a single 4 GB GPU without quantization, distillation, or pruning.
BitNet.cpp now includes VibeASR.cpp, a real-time multilingual ASR inference engine on CPU using BitNet I2_S quantization.
ArtCraft is an IDE for interactive AI image and video creation with 2D composition and 3D scene staging.
The agent skill generates one self-contained HTML + SVG diagram file and supports 45 diagram types.
The browser app shows live aircraft, ships, satellites, earthquakes, traffic, and public cameras on a photorealistic 3D globe.
Open WebUI is a self-hosted AI platform that runs offline with Ollama and OpenAI-compatible APIs.
The project says it offers 61 ToS-friendly providers, 1.3B+ free tokens per month, and support for 11 coding agents.
The same feed, delivered daily at 8 PM ET. No spam, unsubscribe anytime.