LoRA RFT with a helpfulness-weighted composite reward can collapse an adapter's pedagogy mid-training while responsiveness stays intact¶
Claim · All claims
Evidence · 1 study · causalr2·q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly
Subclaims¶
Reading the evidence codes
q is the design tier, i the impact, n the sample, and the kind with its r
the rigour judged within that kind: what each code means.
This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.
q2 i? After LoRA RFT on the Small tier, both adapters land near zero on Axis I yet diverge on helpfulness: metaclaw-rft holds Help flat (4.73→4.76) while openclaw-rl-rft collapses (3.19→2.30), with neither losing responsiveness. → 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? · causal · r2
LoRA RFT (rank-16, 5 epochs) on the Small tier against a composite reward over per-scenario normalized Axes I–III, epoch-averaged (n= 275). "metaclaw-rft holds Help flat (4.73→4.76)" while openclaw-rl-rft's Help falls 3.19→2.30 with no recovery by epoch 3.
"After train- ing both adapters land near zero on Axis I yet diverge on helpfulness, as metaclaw-rft holds Help flat (4.73→4.76) whileopenclaw-rl-rftcollapses,itsHelpfalling3.19→2.30"
Discussion¶
Related Claims¶
- Tutoring quality belongs to the base model and agent harness together rather than either alone, as adapter rankings reorder across tiers — related
- Responsiveness and helpfulness trade off: always-answer adapters score lowest on helpfulness while withholding adapters score highest — related
- Almost no model-and-harness combination sustains tutoring over the full 30-day horizon, as learning plateaus within 5–10 days — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
educlaw-bench-rft-helpfulness-collapse |
title |
LoRA RFT with a helpfulness-weighted composite reward can collapse an adapter's pedagogy mid-training while responsiveness stays intact |
description |
LoRA RFT with a helpfulness-weighted composite reward can collapse an adapter's pedagogy mid-training while responsiveness stays intact |
status |
review |
evidence_strength |
moderate |
newest_evidence |
2026 |
recheck_by |
2029 |
generated |
by process:wiki-ingest on 2026-10-10 |
trust |
unverified — no one has confirmed this page's content |
sources |
1 — lee-2026 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.