AI Paper Atlas

Research map / Continual learning / Continual fine-tuning of LLMs

Just-In-Time Reinforcement Learning: Continual Learning in LLM Agents Without Gradient Updates

Yibo Li, Zijie Lin, Ailin Deng, Xuan Zhang, Yufei He, Shuo Ji, Tri Cao, Bryan Hooi

ICML 2026 · Continual learning › Continual fine-tuning of LLMs · approach: Continual pretraining of language models · 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
update output logits with memory-based advantages instead of appending them to the prompt (ablation)JitRL · WebArena Adminsuccess rate (%): 48.35 -> 52.31Improved
update output logits with memory-based advantages instead of appending them to the prompt (ablation)JitRL · WebArena Redditsuccess rate (%): 54.57 -> 57.64Improved
use JitRL instead of static agentJitRL · WebArena averagesuccess rate (%): 35.63 -> 46.98Improved
use JitRL instead of WebRL fine-tuning on held-out WebArena-LiteJitRL · WebArena-Litefinal success rate (%): 46.06 -> 60.00Improved
use JitRL instead of GRPO training on Zork1JitRL · Jericho Zork1average score: 16.2 -> 53.0Improved
add Reflexion self-reflection on Map siteReflexion · WebArena Mapaverage success rate (%): 30.62 -> 29.38Hurt
use JitRL with DeepSeek-V3.2 backbone instead of StaticJitRL · WebArena Admin (DeepSeek-V3.2)average success rate (%): 32.07 -> 50.65Improved
use JitRL with GPT-5-mini backbone instead of StaticJitRL · WebArena Admin (GPT-5-mini)average success rate (%): 37.07 -> 48.04Improved
restrict memory retrieval to disjoint tasks (unseen-task generalization)JitRL · WebArena Shoppingaverage success rate (%): 23.44 -> 36.98Improved
use JitRL instead of Reflexion on Reddit with DeepSeek-V3.2JitRL · WebArena Reddit (DeepSeek-V3.2)average success rate (%): 55.04 -> 54.42Hurt
set retrieval neighbor count k within 8-14 instead of outside (ablation)JitRL · Library and Zork1scoreImproved
use JitRL instead of WebRL in training costJitRL · WebArenacost (USD): ~$9900 -> $290Improved

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