T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks¶
L1 exact-token and route-replay audit; no 122B MoE training or 300+ turn cloud sandbox.
Paper information¶
| Field | Details |
|---|---|
| Paper | T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks |
| Institution | National University of Singapore / Tencent (first author). |
| First public release | 2026-09-09 |
| Author code | No public author repository found as of 2026-09-14. |
| Local adapter or method key | t1-terminal-rl |
| Local reproduction code | src/auto_research/agent_research/latest_20260914.py |
Paper method and evidence¶
TITO trains on the actual sampled rollout token IDs; turn-boundary repair corrects drift, and R3 replays MoE routes.
flowchart LR
T[Actual rollout token IDs] --> B[Repair turn boundaries]
B --> R[Replay MoE routes]
R --> A[Audit training and rollout alignment]Original paper figure¶
The original figure shows the terminal training flow. From the original paper; copyright remains with the authors.
Core equation¶
The deterministic reference reports token/route state and verifies transitions.
Reported results¶
Paper Terminal-Bench 2.1 improves from 43.8% to 64.0%; train–inference log-probability gap falls from 0.021 to 0.013. Not local or online evidence.
Local reproduction¶
Three-seed metrics use the same public fixtures, budget, and seeds. The deterministic reference exposes decisions and state transitions; tests verify invariants and boundaries. These are not capability or production results.