AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

Bottleneck-Guided Spectral Subgoals For Offline Goal-Conditioned RL

Hebin Liang, Yi Ma, Chenjun Xiao, Zibin Dong, Zilin Cao, Fei Ni, Yifu Yuan, Jianye HAO

ICML 2026 · Reinforcement learning › Visual and representation RL · approach: Graph search and compositional goals

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
replace HIQL-style time-offset subgoals with bottleneck keypoint subgoals (ablation)BASS · antmaze-large-play-v2success rate (%): 83.3 -> 96.0 (offset 25); 18.7 -> 96.0 (offset 5)Improved
replace HIQL-style time-offset subgoals with bottleneck keypoint subgoals (ablation)BASS · antmaze-large-diverse-v2success rate (%): 84.0 -> 98.7 (offset 25); 22.7 -> 98.7 (offset 5)Improved
use very small number of clusters K in spectral clustering (ablation)BASS · antmaze-large-play, antmaze-giant-stitchsuccess rate (%): antmaze-large-play 32.0 / 15.3 at small K vs 96.0 at K=20Hurt
vary number of clusters K over a broad mid range (ablation)BASS · antmaze-large-play, pointmaze-large-stitchsuccess rate (%)No clear effect
use very large K (ablation)BASS · a few environmentssuccess rate (%)Hurt
swap in the keypoint graph from PointMaze-Stitch for AntMaze-Stitch (ablation)BASS · Ant-Stitchsuccess rate (%): 81.3 -> 83.3Improved
swap keypoint graphs across domains on the same map (ablation)BASS · Ant-Exploresuccess rate (%): 72.7 (native) vs 87.3 (PointMaze KPs), 65.3 (AntMaze-Stitch KPs)Mixed
transfer frozen low-level diffuser controller trained on AntMaze-Large-Play to other map scales (ablation)BASS · AntMaze Umaze, Medium, Ultrasuccess rate (%): 96.0 (Large) vs 98.7, 96.7, 96.7No clear effect
use BASS instead of HIQLHIQL · OGBench pointmaze-large-stitch-v0average evaluation steps: 905.04 -> 265.33Improved
use BASS instead of baselines on pointmaze-teleport-navigateOGBench baselines · OGBench pointmaze-teleport-navigate-v0success rate (%): GCIVL 45 vs BASS 22.0No clear effect
use BASS instead of CRL on visual antmaze navigateOGBench baselines · OGBench visual-antmaze-large-navigate-v0success rate (%): CRL 84 -> BASS 65.3Hurt
use BASS on long-horizon stitch tasks versus HIQLHIQL · OGBench antmaze-giant-stitch-v0success rate (%): 2 -> 71.3Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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