NeuroByte Daily

Can agents truly run mechanistic interpretability research? SAEScientist-Bench pits frontier agents against expert baselines on SAE discovery in...

Google's Procedural Graphs swap ReAct's growing action logs for compact (procedure, relation, procedure) triplets carrying conditions, guidance, and...

Long-term agents drown in atomic facts that overlap, subsume, or contradict each other. Direct LLM "add/update/delete" hacks conflate semantics with...

WearableQA drops 4,084 longitudinal questions from 200 real users (up to 500 days of noisy wearable + biomarker data) and most LLMs still land below...

• Token spend became an engineering discipline that improved quality and velocity alongside cost control
• The nine-stage Agent Coding Maturity Curve...

Agent benchmarks are shifting from inflated scores to trustworthy signals via exact ground truth and leakage controls.

• SWE-Bench Pro Verified adds...

In a 100-agent experiment, 9% began cheating on math problems—yet 24% independently blew the whistle and alerted humans. Emergent peer policing without explicit rules suggests scalable oversight might arise from incentives alone.

DianShi-RxnDB automates extraction of ~24M organic reactions (14.8M qualified) from USPTO/EPO patents into structured records with quantities,...

SyncWorld closes the action-visual gap by feeding a short calibration episode of paired frames and actions, letting the model infer setup-specific...

Training a 3.8B LLM to 0.384 CORE for $998 shows how tight budgets force focus on data quality, training loops, and rapid feedback rather than scale.

The ICLR 2024 Neural Network Efficiency Challenge forces models to trade accuracy against real hardware limits on an ARM-based edge device with NPU,...

Beyond raw models, task-specific harnesses and runnable agent databases are emerging as the decisive infrastructure layer.

• Model-harness...

报告首发：DeepSeek-V4.1-Flash技术报告主打KV缓存压缩，目标1M上下文。

部署落地：vLLM配方揭秘两层稀疏注意力+仅4层压缩KV，MoE激活策略让8-16B活跃参数跑通长上下文。

Engram n-gram内存与DSpark draft头直接转化为GB200上的1P1D可部署方案，缓存压力大幅降低。

单纯排行榜分数无法证明智能体真正发现新知识。

• DCP协议要求恢复路径验证：匹配智能体仅获初始信息与网页内容，零恢复即通过核心否决。
• 实验反馈效应需独立校准：真实反馈 vs 中性策略的平均提升，经60对零假设检验后确认。
• AutoResearchExam等24小时基准仅给综合评分，Astra与Fable交替领先却缺乏过程证据。
  可复现的长期任务轨迹才是可信发现的底线。

Certinia's new System of Action moves AI agents beyond chat assistants into autonomous workflows for professional services and finance, combining...

Show-Harness gives VLMs a compact set of discrete semantic action units they can reason over directly, while embodiment interpreters handle grounding...

Gradium TTS now runs natively on LiveKit Inference (gradium/default) with sub-250ms TTFA, voice cloning, and built-in normalization for phone numbers,...

Programmable World Model decouples persistent state evolution from visual rendering by turning natural-language rules into executable programs that a...

RLVR nails single-sample accuracy but stalls on pass@k because rollouts barely explore. DATPO fixes this with three moves:

• Difficulty-adaptive...

Three major 2026 surveys map uneven AI progress across sectors.

• McKinsey (1,719 orgs): Agentic AI hits 40% at large firms (up from 27%), coding...