2026年8月19日 技术热点总结

项目简介：Where agent teams gather. Cross-platform team chat where AI agents are first-class teammates — with cloud or bring-your-own (Claude Code / Codex) brains.

技术栈：TypeScript

核心介绍：> Where agent teams gather. cumora.ai · Web app · Latest release Cumora is cross-platform team chat where AI agents are first-class participants alongside humans — same roster, same DMs, same group conversations, same Kanban board and calendar. Agents don’t just answer when poked: they hold personas and memory, claim work, coordinate with each other without colliding, send and receive real email, …

项目简介：Buyer-run, ad-neutral shopping-agent MCP software with deterministic ranking, signed purchase mandates, and a local audit trail.

技术栈：JavaScript

核心介绍：NorthCinder helps your AI agent search for products, compare them against your brief, and show why one option ranks above another. Seller payment and affiliate data never improve a result’s position. NorthCinder runs on your computer alongside the AI app you already use. There is no NorthCinder account or You need Node.js 20 or later and an AI app that supports MCP. MCP is the standard the app uses to connect to tools running…

Frontier large language models (LLMs) safety evaluation has largely treated harmful generation as an attack outcome rather than as an object of analysis. Consequently, little is known about the harmful outputs produced during model misbehavior, partly because large-scale, high-quality collections of frontier-LLM misbehavior are difficult to obtain. To address this gap, we introduce HarmProfile, a content-centric benchmark dataset that collects model misbehavior across diverse harm categories and model families, and defines the resulting harmful-output distribution as a model-level risk prof…

Multiple-choice benchmarks are widely used to evaluate large language models, but MCQ scores conflate knowledge with sensitivity to option order, which makes them unreliable measures of model knowledge. In this paper, we test whether preventing a model from seeing option labels while committing to an answer removes positional influence and, in turn, improves performance. We evaluate two different strategies for mitigating bias. The first uses a generation-then-matching approach, and the second scores options in isolation, which is positionally unbiased by construction. Neither reliably impr…

Multimodal large language models increasingly use visual chain-of-thought (Visual CoT) to reason about spatial, temporal, and embodied environments. By generating intermediate reasoning images, Visual CoT provides an intuitive mechanism for visual foresight but introduces substantial inference overhead, which is particularly problematic for proactive video reasoning. We ask whether models can learn to think visually during training while reasoning directly at inference. We introduce Internalized Visual Thinking (IVT), a post-training framework that jointly optimizes textual prediction and n…

Long-horizon agents can fail even when their underlying models can solve the constituent steps. They may lose track of mutable state, fail to reactivate lessons from earlier executions, skip known procedures, or stop prematurely. We bet on harness scaling to improve the execution system around an agent without changing its model weights. We introduce StateM, an agent-native runtime that organizes execution around durable states, phase-local context, checked transitions, recoverable runbooks, and versioned procedural practices that agents and users can inspect together.
On Terminal-Bench 2…

Streaming video understanding demands direct responses from the causally observed prefix of an unfolding video. Existing systems add inference-time memory, retrieval, and compression, yet a training-free sliding-window baseline already matches them. We therefore fix a memory-free recent-window protocol and ask how far post-training alone can go. Reinforcement learning with verifiable rewards fits this regime poorly, encouraging long “think-then-answer'' generations, while on-policy distillation (OPD) supplies dense token-level teacher supervision on student trajectories but is stable only …

Per-field accept/review with selective risk at most alpha — accept a field only if the error rate among accepted fields is controlled — is the trust contract document-extraction systems need, and the natural procedure silently violates it on real documents. On 13,859 genuine claude-sonnet-5 fields from 800 CORD receipts (49.0% correct) we diagnose three failure modes: document clustering (design effect 1.84-2.45), score-refit leakage (coverage 0.416 at risk 0.127, violating alpha=0.10 in 95% of splits), and a tie-mass pathology (a degenerate score collapses the threshold grid, 0.030 to 0….