AI Paper Atlas

Research map / Large language models / Agents, code and math

Grounded in Reality: Learning and Deploying Proactive LLM from Offline Logs

Fei Wei, Daoyuan Chen, Ce Wang, Yilun Huang, Yushuo Chen, Xuchen Pan, Yaliang Li, Bolin Ding

ICML 2026 · Large language models › Agents, code and math · approach: Long-horizon LLM agents · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
train with Learn-to-Ask (GRPO) instead of base modelQwen2.5-7B-Instruct · RealMedConv evalWA-GH: 0.13 -> 0.41Improved
train with Learn-to-Ask (GRPO) instead of base modelQwen2.5-7B-Instruct · RealMedConv evalWS: 0.16 -> 0.93Improved
train with Learn-to-Ask (GRPO) instead of base modelQwen2.5-32B-Instruct · RealMedConv evalWA-GH: 0.13 -> 0.37Improved
add question utility reward Ra (ablation)Learn-to-Ask 7B · RealMedConv evalWA-GH: 0.19 -> 0.41Improved
add assessment accuracy reward Rs (ablation)Learn-to-Ask 7B · RealMedConv evalWS: 0.02 -> 0.93Improved
multiplicative reward fusion instead of summation (ablation)Learn-to-Ask 7B · RealMedConv evalTR: 3.20 -> 3.27Improved
multiplicative reward fusion instead of summation (ablation)Learn-to-Ask 32B · RealMedConv evalTR: 3.19 -> 3.15No clear effect
use CISPO instead of GRPO (ablation)Learn-to-Ask 7B · RealMedConv evalTR: 3.27 -> 3.36Improved
use GSPO instead of GRPO (ablation)Learn-to-Ask 7B · RealMedConv evalTR: 3.27 -> 3.16Hurt
increase beta from 2 to 4 (ablation)Learn-to-Ask 7B · RealMedConv evalTR: 3.27 -> 3.15Hurt
SFT instead of Learn-to-AskQwen2.5-7B-Instruct · RealMedConv evalWA: 0.50 -> 0.40Hurt
DPO instead of Learn-to-AskQwen2.5-32B-Instruct · RealMedConv evalWA: 0.50 -> 0.23Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy