TMax Co-Author to Debate Data Recipes for Terminal Agents at RL Paper Club
Shahules786 · x · 2026-11-03
Vibrant Labs' RL Paper Club hosts Hamish Ivison, co-first author of TMax, comparing three recipes for training terminal agents:
- TMax (open baseline): 14,600 synthetic terminal environments sampled across nine axes, each turned into a Dockerfile, verifier, and instructions; outcome-only RL takes Qwen 3.5 9B to 27% on Terminal-Bench 2.0. It only checks that images build, letting RL filter broken tasks.
- Terminal-World: builds environments from human-authored agent skills, deriving task, sandbox, verifier, and teacher guideline from one skill, validating every artifact, trained by SFT alone.
- T1: the large-scale RL end—a 122B MoE trained with PPO for 300+ tool-call turns, rewarded by the number of verifier checks passed rather than pass/fail.
Debate topics: where tasks come from, whether RL filtering can substitute for validation, reward granularity, SFT vs RL, and whether a synthetic recipe can beat the model that generated it.
More from coding & agent
- Agent-built data pipeline flags law firms whose clients later hit SEC/DOJ trouble — doodlestein · 2026-11-03
- MIT-licensed frontend skill gives coding agents a 7-step design and verification workflow — Extra-Highlight9183 · 2026-11-03
- Trajectory launches post-training platform with GS1 and Continual Graders research previews — simonguozirui · 2026-11-03
- SMBs are rebuilding their entire GTM stack with agent-native tools, says new thread — garrytan · 2026-11-03
- Compacting Claude Code at 400K instead of 967K cut this dev's cost per request by 43% — Awkward-Let-4628 · 2026-11-03
- Capy 0.7.0 ships Captains orchestration GA; YC's Garry Tan reports 5X issue-fixing throughput — garrytan · 2026-11-03