Research map / Reinforcement learning / Visual and representation RL
Path Channels and Plan Extension Kernels: a Mechanistic Description of Planning in a Sokoban RNN
Mohammad Taufeeque, Aaron David Tucker, Adam Gleave, Adrià Garriga-Alonso
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 changed | Where | Measured | Effect |
|---|---|---|---|
| intervene on 59 path channels via single-step cache ablation (ablation) | DRC(3,3) · Boxoban medium | solve rate drop (%): 57.6 ± 2.8 | Improved |
| intervene on 37 non-path channels (ablation) | DRC(3,3) · Boxoban medium | solve rate drop (%): 10.5 ± 1.9 | No clear effect |
| intervene on random 37 path channels (count-matched) (ablation) | DRC(3,3) · Boxoban medium | solve rate drop (%): 41.3 ± 2.4 | Improved |
| edit pooled next action channels causally | DRC(3,3) · 10,000 Boxoban transitions | intervention score (%): 99.7 ± 0.2 | Improved |
| use channel-based intervention instead of linear probe (box movement) | DRC(3,3) · 10,000 Boxoban transitions | intervention score (%): 82.5 ± 2.5 -> 86.3 ± 2.1 | Improved |
| use agent-movement channel intervention instead of probe | DRC(3,3) · 10,000 Boxoban transitions | intervention score (%): 20.7 ± 0.7 -> 53.2 ± 2.1 | Improved |
| mean-ablate cell state c (ablation) | DRC(3,3) · Boxoban medium | solve rate drop (%): 21.28% ± 0.04% | Improved |
| mean-ablate forget gate f (ablation) | DRC(3,3) · Boxoban medium | solve rate drop (%): 2.66% ± 0.03% | No clear effect |
| mean-ablate i, j, o gates or hidden state h (ablation) | DRC(3,3) · Boxoban medium | solve rate drop (%): 100.00% | Improved |
| zero-ablate kernels between box-down and box-right channels (remove winner-takes-all) (ablation) | DRC(3,3) · two equally viable paths level | plan selection | Improved |
| scale up plan extension kernels by factor 1.4 | DRC(3,3) · 40x40 level | solving long paths | Improved |
| steer W_ch1 and W_ch2 kernels by factor 1.2 | DRC(3,3) · 16x16 zig-zag level | solving level | Improved |
Closest papers by meaning
- Efficient Reinforcement Learning by Discovering Neural Pathways (NeurIPS 2024 · similarity 0.86)
- Differentiable Learning of Lifted Action Schemas for Classical Planning (NeurIPS 2026 · similarity 0.86)
- ALPINE: Unveiling The Planning Capability of Autoregressive Learning in Language Models (NeurIPS 2024 · similarity 0.85)
- Interpreting Emergent Planning in Model-Free Reinforcement Learning (ICLR 2025 · similarity 0.85)
- Graph Neural Network Based Action Ranking for Planning (NeurIPS 2025 · similarity 0.85)
- Amortized Planning with Large-Scale Transformers: A Case Study on Chess (NeurIPS 2024 · similarity 0.85)
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.
