Frontier Theory Radar：按决策价值筛选前沿 AI 论文的每日研究系统

Evidence-first decisions on which frontier AI papers matter now, later, or not yet.

A daily research system that separates immediate engineering value, emerging trends, durable long-tail ideas, and noise.

简体中文 · Live radar · Daily reports · Methodology

Latest research run · 2026-10-10

Papers reviewed	Immediate	Trend	Long tail	Deferred
10	132	220	233	1335

Deep dive: CSF: Contextual Safety Filtering for Motion Generators · Immediate · 轻量试点

Recent reports

Date	Deep dive	Value	Decision
2026-10-10	CSF: Contextual Safety Filtering for Motion Generators	Immediate	轻量试点
2026-10-08	Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts?	Immediate	轻量试点
2026-10-07	Base Models Can Reason By Taking a Cue From Training Data	Immediate	轻量试点
2026-10-06	LESSER: Post-Training Data Selection with Output-Layer Gradients	Immediate	轻量试点
2026-10-05	VISTA: A Visual Harness for Reasoning in an Interactive World	Immediate	轻量试点
2026-10-04	VISTA: A Visual Harness for Reasoning in an Interactive World	Immediate	轻量试点
2026-10-03	KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards	Long tail	持续观察

Directions under active observation

Direction	Stage	Related papers
Agentic World Modeling	上升	596
Coding Agent	主流化	584
Context Engineering	上升	596

Why this repository exists

Paper feeds optimize for recency and popularity. Frontier Theory Radar optimizes for decisions: what deserves an experiment today, what needs repeated observation, what should be preserved for later, and what does not yet justify attention. Each conclusion keeps its source link and states the largest remaining uncertainty.

Research workflow

flowchart LR
  A["Collect papers"] --> B["Score relevance and evidence"]
  B --> C["Route by value"]
  C --> D["Deep dive"]
  C --> E["Trend watch"]
  C --> F["Long-tail archive"]
  D --> G["Action and reusable asset"]

papers/ stores dated source snapshots and scored candidates.
daily/ stores the human-readable decision record for each run.
trends/ and insights/ retain multi-day directions and reusable findings.
docs/data/ is the structured public projection used by the live site.
scripts/generate_readme.py rebuilds both READMEs and the activity chart from committed evidence.

Evidence boundaries

Scores and decisions are research judgments, not claims of independent reproduction. A paper abstract, author-reported benchmark, open-source implementation, and independently reproduced result are treated as different evidence levels. Missing code, truncated abstracts, or unverified benchmarks are called out rather than silently upgraded into facts.

Run and verify

./run_daily.sh 2026-08-07
python3 -m pytest tests
python3 scripts/generate_readme.py

The scheduled production job runs in the private AIUtil automation environment. Credentials and private runtime memory are not stored here. The generated Markdown, SVG, JSON, and source-linked reports remain reviewable in Git history.

Security

Do not commit feed credentials, API tokens, private paper collections, or operator memory. Report a vulnerability privately through GitHub Security Advisories.

License

Apache License 2.0. See NOTICE.