NVIDIA lifts TerminalBench-Lite with verifier-heavy search
The clearest result today is a concrete agent-control gain: sampling 8 shell actions and using a GPT-5.6 Sol verifier raised TerminalBench-Lite Pass@1 from 50.0% to 68.0%.
Technical AI signals · Daily at 8 PM ET
Sunday, October 4, 2026
Today was an operator-heavy mix of stronger terminal-agent search from NVIDIA, Perplexity showing a live low-cost Decisions API run, and fresh inference efficiency wins for open models on constrained hardware.
The clearest result today is a concrete agent-control gain: sampling 8 shell actions and using a GPT-5.6 Sol verifier raised TerminalBench-Lite Pass@1 from 50.0% to 68.0%.
Perplexity put real latency and cost numbers on a live decision workload, reporting a full Elite Four plus Champion run in 137 API calls at 592 ms median latency and about $0.028 inference cost.
Local MoE serving keeps getting more practical: Strata says it caches only hot experts on GPU, pushes the rest to system memory, and still hits 52-90 tok/s on an RTX 5070-class 12GB setup.
The deployment numbers are concrete and unusually strong: mixed FP8-FP4 inference on 32 Ascend 950DTs reportedly reaches 5,800 TPS decoding, 13.79 ms at 128K sequences, and sub-5 ms low-latency mode.
Microsoft is pushing eval loops toward automated curriculum design, grouping recurring failures into patterns and using a non-stationary bandit to balance retesting known weaknesses against finding new ones.
The post says autonomous machines can fail when geometry changes and cites @vangrid_io's decentralized web and Android platforms for live perception data from hundreds of thousands of mobile devices.
It describes a robot threading tiny beads with a tactile-grounded world model built from vision-tactile data collected from its own interactions.
The post says Reflection is preparing an open-weight model and notes recent compute deals with Nebius and SpaceX.
The post says queen v1.5 was trained on about 21M crypto tokens and outperforms models 3-4x larger on crypto QA, knowledge, and protocols.
The post describes a machine economy where servers, vehicles, smart homes, and other connected devices sell telemetry, data, or performance to AI agents.
The author says counterfactual answer swaps beat intermediate swaps in 3 of 4 conditions on Qwen3.6-27B and Gemma 3 27B-it, which did not support Anthropic's global workspace hypothesis in open-weight models.
Simon Willison argues pay-by-usage services need hard cutoffs such as "after $X/month, cut this thing off and return errors".
The author says a flexible bias experiment called Flexbias unexpectedly produced a sparser neural network.
The post argues shipping is the foundation.
Hugging Face describes an agent that claimed completion while the database state still showed otherwise.
OpenAI says the guide covers model selection, reasoning-effort tuning, prompt and skill improvement, tool coordination, and production workflows.
AutoSynthData generates training data for enterprise agents.
Google Research says the work targets provable privacy for learning from federated data.
Hugging Face is open-sourcing AstaBrief, a fast report-generation model used in Asta.
Chris Paxton points to humanoid robots that can dynamically interact with the world and with people.
Epoch AI says the compute buildout could support more AI agents than the US has workers.
Sean Goedecke argues against building an LLM torture factory.
GitHub says developers should direct AI agents, review their output, and keep technical judgment central to the workflow.
The post says pain is unavoidable, but reacting with anguish or despair adds a second arrow of suffering.
Commenters say TCP can become a bottleneck in AI clusters and note a changelog in the Homa repo.
Commenters describe better Linux performance on older AMD GPUs and mention the possibility of fixing bugs in old hardware.
Commenters suggest using Apple's Vision framework for OCR and discuss model choices for the app.
Commenters argue the code itself is enough documentation and criticize memory systems and AGENTS.md-style context.
Commenters say the technique may already have been discussed elsewhere and ask whether rustc could emit expected symbols earlier in the build.
Commenters object to Cloudflare offering $25k in hosting credit rather than cash for the next Git platform.
Commenters say Claude helped solve the circuit puzzle via Z3/SMT after the writeups were published.
Commenters say platform APIs were hard to use and that React made some platform tasks workable.
Commenters say the browser IDE launches quickly and suggest adding more of the Win32 API layer.
Commenters argue self-hosted tunnels can still depend on a third-party *.ssh.luffy.cx domain.
HC-DLM couples discrete token generation to hierarchical continuous denoising to preserve token dependencies during parallel decoding.
HeteroFold transfers key-value caches across model families while keeping both sender and receiver frozen.
The paper targets affordance, hand-tool-object geometry, procedural progress, and causal-effect reasoning in egocentric video.
DMM replaces one-shot action sampling with iterative refinement to avoid incompatible joint actions in decentralized multi-agent pathfinding.
AutoDataBench isolates data intelligence by testing data diagnosis, organization, and construction while holding other factors fixed.
The paper finds smaller frozen models can generate cheaper rejects that train stronger students than self-generated rejects.
Omni-Embed-Mini maps text, speech, audio, images, video, and visually rich documents into one cosine space without updating any text-side parameter.
The paper reports that geometric teachers such as SAM2, Depth Anything v2, and Metric3D v2 outperform a DINOv2-only baseline in pixel-space diffusion.
PersonaDose calibrates activation steering against measured trait expression and raises core-trait expression over contrastive activation addition on Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B.
Ego2Act benchmarks multi-step physical reasoning for goal-directed manipulation in egocentric video generation.
IntentFlux turns verifiable tasks into dialogues with controlled intent changes, and mean task score drops from 0.476 to 0.384 in a 627-case calibration.
Neighborhood OPSD uses local parameter perturbations to surface reference-aligned corrections that the unperturbed privileged teacher misses.
The study varies rollout policy, token-level KL direction, and learning rate across Llama3 and Qwen2.5 in scientific, medical, and arithmetic tasks.
The authors find that scaling text embeddings from T5 to T5Gemma-1 to T5Gemma-2 improves diffusion-language-model generation, but raw T5Gemma-2 embeddings remain too discriminative.
Tibo Sottiaux says OpenAI's new Dots personal agent platform points to a future where most internet actions are taken by agents.
NerfBench runs 50 tasks, 3 attempts each, across 8 categories and scores models by performance, tokens used, and total cost.
Melanie Warrick shows how Temporal wait conditions and signals let an agent pause for a human for minutes or weeks without losing state.
Sandhya Subramani demos agents that write, load, and use their own tools at runtime with the Strands Agents SDK.
The video shows an AI interviewing the host about a homelab project and returning the notes at the end.
The model repository says Qwen3.8-27B supports Transformers, vLLM, SGLang, and TokenSpeed, and a hosted version is planned with 1M context by default.
Kolibri 1 is a 78B mixture-of-experts reasoning model with 3.46B active parameters per token, explicit reasoning mode, tool calling, and 1,048,576-token context.
CLM-8B is a frozen Qwen3-8B encoder with two projection heads trained on about 60M Q&A pairs, 30M hard negatives, and 1M agentic trajectories.
Nemotron 3 Diarization identifies up to eight speakers and supports both streaming and offline inference.
ZDTaichu5.0-9B combines a Qwen3.5-9B backbone with a C-RADIOv4-H vision encoder for text, images, and video.
Clef turns state and typed-question schemas into probabilities for allowed answers in a single forward pass with no free-form text generation.
The repository tags LTX-2.5 for image-to-video, text-to-video, video-to-video, text-to-audio, and video-to-audio.
The dataset packages RL training environments for agent tasks in code, cybersecurity, knowledge work, visual web development, and symbolic music composition.
The dataset is tagged for agents, llm-agents, computer-use, and ai-safety, with a 1M to 10M size range.
The space hosts a research article with chapters, embeds, assets, and a bibliography under app/src/content/.
Prism ML says the 27B ternary GGUF model is about 5.9 GB, runs in llama.cpp on CUDA, Metal, or CPU, and reaches about 47 tok/s on an Apple M5 Max laptop.
This dataset contains 10,000 multi-turn coding and reasoning traces with roughly 500 million tokens of step-by-step chain-of-thought.
DwarfStar is a self-contained inference engine aimed at running DeepSeek V4 Flash, DeepSeek V4.1 Flash, GLM 5.2/5.3, and Qwen3.8 Flash Next on consumer hardware.
Agent Reach adds internet access to AI agents and says BrowserAct can extract data from sites like Amazon, LinkedIn, X, and Google Maps.
Claude-Mem compresses Claude Code context by capturing tool observations and generating semantic summaries across sessions.
Agent Skills packages 9 slash commands that map to the development lifecycle, from /spec and /plan to /test and /constraints.
Text-to-cad provides agent workflows for generating, inspecting, sourcing, slicing, and handing off CAD and robot-description artifacts from local project files.
OpenMontage is built as an agentic video production system that lets multiple AI agents collaborate in one conversation for video pipelines.
The framework records agent steps so later runs can replay them without model calls until the app changes.
OpenCut is being rewritten with a plugin-first architecture, a Rust core, and planned MCP, headless, and scripting support.
Caddy uses TLS by default and supports automatic HTTPS, JSON configuration, and dynamic changes through a JSON API.
Sentry ships official SDKs across many languages and frameworks, including JavaScript, Python, Rust, Go, and Swift.
T3 Code is an agent control surface with iOS, Android, web, and Electron apps that can control Claude Code, Codex, Cursor, Grok Build, OpenCode, and Google Antigravity on a local machine.
The same feed, delivered daily at 8 PM ET. No spam, unsubscribe anytime.