Direction, evidence grade, and study type are AI-generated labels (gpt-5-mini), not human-verified. Syntheses are LLM-written. "Tensions" are machine-detected candidates, not confirmed contradictions. A research-acceleration tool, not peer review.
How this is built →
1Distinct papers
1Unique collaborators
1/1Semantic Scholar citation coverage
Publication span: 2026. Corpus fetch span: 2026.
Identity provenance
Provider IDs
- Semantic Scholar:
2458353771
ORCID evidence
No valid ORCID is stored.
Observed aliases (1)
- Alexander Tu (semantic scholar, provider refresh)
Topics and outcomes in this view
Assessment themes
- Governance: 1 paper
- Human Ai Collab: 1 paper
Claim outcomes
- Ai Safety And Ethics: 1 paper
- Task Completion Time: 1 paper
- Governance And Regulation: 1 paper
Papers in the Semantic Scholar view
Latest stored Semantic Scholar author observations only. Citation counts below are from the same provider and are not combined with other services.
Scroll the table horizontally to see every column.
| Paper | Author evidence | Date | Provider citations |
|---|---|---|---|
| Fine-tuning a 4B language-agent to prefer minimally sufficient privileges cuts unsafe or over-privileged actions from ~4.6% to ~0.8% and raises safe task completion from ~64% to ~98% across a 500-task held-out synthetic evaluation, showing learned restraint can complement permission gates.arxiv | Alexander Tu provider id |
2026-08-18 | 0 |
Citation observation summary
Semantic Scholar supplied counts for 1 of 1 papers in this view; 0 are missing. The observed paper counts sum to 0 cumulative citations. This is a coverage summary, not an author score or h-index.