Responsiveness and helpfulness trade off: always-answer adapters score lowest on helpfulness while withholding adapters score highest¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? On the frontier tiers, the always-answer adapters (picoclaw, deeptutor, ironclaw) score lowest on Helpfulness (3.8–4.6) while the withholding adapters (openclaw, zeroclaw) score highest (5.6–6.1), with direct leakage staying negligible (0.05% hand-over). → Lee 2026

Evidence¶

Lee 2026¶

Lee, U.; Lee, S.; Jeong, Y.; Lee, E.; Shin, M.; Kwon, H. (2026). EduClaw-Bench: A Long-Horizon Benchmark for Pedagogical LLM Agents with Simulated Learners. https://arxiv.org/abs/2608.03206 doi:10.48550/arxiv.2608.03206

q2 · i? · design · r2

Benchmark results on frontier tiers showing the inverse relationship between responsiveness (fraction of help-requests answered) and helpfulness (29-item LearnLM rubric). Always-answer adapters reach near-100% responsiveness but score 1.5–2 points lower on helpfulness.

"Responsiveness spans25–100%and runs inversely to Helpfulness on the frontier tiers, where the always-answer adapters (picoclaw, deeptutor, ironclaw) scorelowestonHelp(3.8–4.6)andthewithholdingadapters (openclaw,zeroclaw)highest(5.6–6.1),whiledirectleakage staysnegligible(0.05%hand-over)."

Discussion¶

Related Claims¶

• LoRA RFT with a helpfulness-weighted composite reward can collapse an adapter's pedagogy mid-training while responsiveness stays intact — related
• Using the directness penalty Rdir as a GRPO reward is counterproductive: it lowers SE by 0.200 and raises leakage by 10 percentage points relative to omitting it — related