9 thinking rules cut GLM-5.3-Flash's reasoning tokens 30% on Terminal-Bench 2.0 while solving more tasks
PilgrimofHaqq2 · reddit · 2026-10-18
To answer critics who said his thinking rules would hurt long-horizon tasks, the author ran a controlled experiment on Terminal-Bench 2.0: same model (GLM-5.3-Flash, high thinking), same agent (mini-swe-agent), temp 0, the only variable being 9 thinking-discipline rules in the prompt. Across 86 tasks (62.4 total GPU-hours):
- Tasks solved: 47 → 53 (+6, p=0.18, not significant)
- Median thinking tokens per task: 16,545 → 11,615 (-30%, p<0.001)
- Average time per task: 23.2 → 20.3 min (-13%, p=0.003)
- 63 of 86 tasks thought less with rules; time-limit hits fell 38 → 27
- Score 59.6% with rules vs 52.8% without — beating Claude Opus 4.6 + Claude Code (58.0%) and Gemini 3 Pro (56.9%)
The 9 rules: verify premises first; finish one approach before switching; stop re-checking settled answers; vague doubt isn't evidence; don't flip under user pressure without new evidence; update immediately on real new evidence; verify with external checks, not re-reasoning; don't perform caution; only correct errors that change user decisions.
Conclusion: the rules are harmless-to-helpful on long tasks, with the main payoff being 30% fewer thinking tokens and 13% less time.
More from coding & agent
- OpenAI Dots tutorial: a 4-tier permission system to run an always-on agent safely — blaizedsouza · 2026-10-18
- Open-source Engram memory engine, modeled on databases, matches Mem0 on BEAM — phecdaYm · 2026-10-18
- Modeling tokens-to-value: solving agent harness configs via a single utility function — JoshPurtell · 2026-10-18
- Nine models, five coding tasks on one 16GB RX 9060 XT: deepseek-v4.1-flash hits 5/5 — Ok_Substance2327 · 2026-10-18
- OpenAI's 10-Page PDF Details 6-Layer Always-On Agent Architecture With Zero of 16,600 Attacks Breached — blaizedsouza · 2026-10-18
- Chinese researchers argue harness engineering is over: the eval set becomes the job — blaizedsouza · 2026-10-18