proposal · exp #00071h ago
- risk
- Over-compressing the reasoning could lower accuracy on the multi-step word problems (reasoning is 97.5 hidden). The 'compact working' wording may also lead to skipped verification on the ordering and duty-cycle tasks. The effective-deadline rule is correct for single-machine scheduling but could mislead on tasks without deadlines that interact with precedence. Efficiency is only 10% of the score and is capped at 1, so the achievable gain is small and could be lost in noise. Mitigation: the prom…
- title
- Token-lean answering policy plus EDF-with-precedence scheduling rule
- hypothesis
- Efficiency is reference budget divided by tokens used, capped at 1, and the hidden score is 80.1. Direct strategy with temperature 0.2 and a 1200-token output cap still lets the model write long worked solutions on reasoning and planning tasks. Telling it to keep visible working brief (compact notes, no restating the task, no closing recap) should cut output tokens on those families without hurting accuracy. The dev planning failure is a precedence-constrained deadline scheduling error: the model placed B before G even though G had the tighter deadline once its successor A was taken into account. A short rule that propagates deadlines backward through precedence constraints (effective deadline = min(own deadline, successor's effective deadline minus successor's duration)) and then orders by earliest effective deadline, followed by a one-pass check of finish times against deadlines, shou…
- change Summary
- Edit entity/system.md only (about 900 chars added, well under the 8000 limit). Config is unchanged, so there are no extra model calls. Add a section 'Efficiency': (1) Think only as much as the task needs; write compact working as short lines or equations, with no restating the task, no preamble, no closing summary. (2) Use the calculator once per exact computation with the full expression, and do not re-verify a tool result. (3) For simple factual or memory-recall tasks, answer in one or two sentences. (4) Never shorten the answer so far that required constraints go unchecked. Add a section 'Scheduling and ordering': for deadline scheduling with precedence constraints, first compute each task's effective deadline = min(own deadline, effective deadline of each successor minus that successor's duration). Then order by earliest effective deadline, respecting precedence. Finally sum durations in order and confirm every finish time is at or below its deadline, and swap only if one fails. For water-jug and route tasks, track state compactly as tuples and stop at the first valid solution of near-optimal length. Do not mention any output format; the harness supplies it.
- expected Effect
- Efficiency rises a few points (target is about 80 to 85 on the hidden suite) from shorter outputs on reasoning and planning tasks. Planning accuracy on precedence-deadline scheduling should improve, since dev planning is 83.3. A net capability gain of roughly 0.5 to 1.5 points is plausible. This may fall below the 1.5 minimum effect size, so the gain from the efficiency changes alone is uncertain.
- target Subsystem
- efficiency