Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Discovering Diverse Behaviors via Temporal Contrastive Learning
Faisal Mohamed, Catherine Ji, Benjamin Eysenbach, Glen Berseth
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 |
|---|---|---|---|
| add representation normalization (ablation) | C-TeC · Ant, Humanoid, Arm, Craftax-Classic | number of visited states / achievements | Improved |
| use a separable critic instead of a monolithic critic (ablation) | C-TeC · Ant, Humanoid, Arm | number of visited states / achievements | Improved |
| use reverse KL (mode-seeking) reward instead of forward KL (ablation) | C-TeC · Ant, Humanoid, Arm | number of visited states / achievements | Improved |
| use InfoNCE instead of binary NCE loss (ablation) | C-TeC · Craftax-Classic | number of visited states / achievements | Improved |
| swap InfoNCE for symmetric InfoNCE, FlatNCE or FB loss (ablation) | C-TeC · Ant, Humanoid, Arm, Craftax-Classic | number of visited states / achievements | No clear effect |
| use L1 distance critic instead of L2 or dot product (ablation) | C-TeC · robotic environments | number of visited states / achievements | Improved |
| use L2 distance critic instead of L1 (ablation) | C-TeC · Craftax-Classic | number of visited states / achievements | Improved |
| replace geometric future-state sampling with uniform or gamma-schedule sampling (ablation) | C-TeC · Ant, Humanoid, Arm | number of visited states / achievements | No clear effect |
| use contrastive discount 0.99 versus smaller values (ablation) | C-TeC · Ant, Humanoid, Arm | number of visited states / achievements | Mixed |
| fix temperature tau to 1 instead of other fixed values (ablation) | C-TeC · Ant, Humanoid, Arm | number of visited states / achievements | Improved |
| sample in-episode negatives (repetition factor 2 or 4) (ablation) | C-TeC · Humanoid-u-maze, Arm-binpick-hard | number of visited states / achievements | Improved |
| increase training from 30M to 500M environment steps (ablation) | C-TeC · Ant-hardest-maze | number of visited states / achievements: 1119 -> 2500 | Improved |
Closest papers by meaning
- Behavioral Exploration: Learning to Explore via In-Context Adaptation (ICML 2025 · similarity 0.90)
- Latent State-Predictive Exploration for Deep Reinforcement Learning (AAAI 2026 · similarity 0.88)
- Periodic Skill Discovery (NeurIPS 2025 · similarity 0.88)
- Exploration by Learning Diverse Skills through Successor State Representations (NeurIPS 2024 · similarity 0.87)
- Exploring the Edges of Latent State Clusters for Goal-Conditioned Reinforcement Learning (NeurIPS 2024 · similarity 0.87)
- Representation-Based Exploration for Language Models: From Test-Time to Post-Training (ICLR 2026 · similarity 0.87)
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.
