AI Paper Atlas

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

SkillGen: Learning Domain Skills for In-Context Sequential Decision Making

Ruomeng Ding, Wei Cheng, Minglai Shao, Chen Zhao

AAAI 2026 · Large language models › Agents, code and math · approach: Agent instruction-following benchmarks

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
use SkillGen skill prompts instead of Synapse 3-shot expert trajectoriesQwen2.5-7B-Instruct · ALFWorldSuccess Rate (%): 19.4 -> 55.2Improved
use SkillGen instead of Synapse 3-shotGPT-4o-mini · ALFWorldSuccess Rate (%): 17.8 -> 29.8Improved
use SkillGen instead of baselinesQwen2.5-7B-Instruct · BabyAIGrounding Rate (%): 78.6 vs 66.7Mixed
use focused golden segment instead of a fixed or retrieved full demonstration (ablation)SkillGen · ALFWorld (Qwen-Turbo)Success Rate: +17.2% over Fixed, +3.7% over RetrievedImproved
add step-wise skills to the golden segment (ablation)SkillGen · ALFWorldSuccess Rate: +8.1%Improved
add step-wise retrieval instead of injecting all domain skills (ablation)SkillGen · ScienceWorldSuccess Rate: +5.7%Improved
inject all domain skills without retrieval (ablation)SkillGen · ScienceWorldProgress Rate: +3.6% PRImproved
add TD-based credit assignment over segment retrieval with sparse progress signal (ablation)SkillGen (Qwen2.5-7B-Instruct) · ALFWorldSuccess Rate: +12.8 SRImproved
add TD-based credit assignment over segment retrieval (ablation)SkillGen (GPT-4o-mini) · ScienceWorldSuccess Rate: 16.8% increaseImproved
use subgoal annotations as weak supervision rather than only fully successful trajectories (ablation)SkillGen (GPT-4o-mini) · ScienceWorldSuccess Rate (%): 32.4 -> 40.2Improved
increase sampled trajectories per task from 6 to 9 or 12 (ablation)SkillGen · ALFWorld, ScienceWorldSuccess RateNo clear effect
retrieve more distinct actions (s=3, k=1) instead of s=1 (ablation)SkillGen (Qwen2.5-7B-Instruct) · ScienceWorldSuccess Rate (%): 14.4% SRHurt

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