Reward Shaping Ablation: 0.1 Bonus Cuts Tool Calls 11-17% at No Accuracy Cost
yb2698 · x · 2026-11-06
The author details their agent RL reward: verifier score + 0.1 × (1 − calls/15), applied only to correct rollouts. Ablations show a 0.3 coefficient hurts accuracy, while at 0.1 accuracy holds and tool calls drop 11–17%.
Training streams live on a Space showing every safeagents event, train reward, held-out pass@1 per checkpoint, and tool calls per solved task; credentials are redacted before leaving the trainer. A 50-step ablation (pi, LFM2.5-24B) found the guard changes held-out accuracy by +0.0 ± 4.1 points — a free safety measure.
More from coding & agent
- One plain-English prompt turns your phone into a 200-particle physics simulator — tak3sh8 · 2026-11-06
- Zep Agent Skills turns verified agent traces into versioned, governed procedural memory — TheTuringPost · 2026-11-06
- Zep launches Agent Skills: compile verified agent runs into reusable skills — TheTuringPost · 2026-11-06
- Sentry's next MCP evolution: one CLI login powers both MCP and full CLI access — zeeg · 2026-11-06
- Box launches Box Content Retrieval with agentic platform and MCP support at BoxWorks — shashib · 2026-11-06
- Give your agent an interpreter, not tools: 19 min to 2 min on the same model — Itz_hany · 2026-11-06