- llm
- nlp
- reasoning
- rl
- rlhf
- ssm
- theory
- information-theory
- representation-learning
- optimization
- rag
- alignment
- paper-review
- survey
- notes
•
•
•
•
•
•
•
•
•
•
•
•
•
•
-
SOL — Self-Optimizing Language Models via Token-Level Efficiency Policies
Paper review of SOL (Akhauri, Abdelfattah, ICML 2026, Cornell) — a framework that trains a lightweight policy network (via GRPO) to assign per-token efficiency actions (attention sparsity, MLP pruning, quantization) to a frozen LLM, allocating more compute to important tokens. A joint SOL-J policy achieves +7.3% MMLU accuracy over uniform budget baselines, consistently improving the Pareto front across Llama and DeepSeek-R1 models.
-
WFM-TTS — Test-Time Scaling for World Foundation Models
Paper review of "Can Test-Time Scaling Improve World Foundation Model?" (Cong, Zhu et al., COLM 2025) — the first framework that applies test-time compute scaling to video-generative world models. A 9,000x-faster efficient decoder enables beam search + top-K pruning over the token space, rule-based rewards guide selection, and a 4B COSMOS model with WFM-TTS matches a 12B baseline. Introduces a new five-dimension WFM evaluation toolkit (3D consistency, temporal coherence, spatial awareness, perceptual quality, text-video alignment).
-
Monte Carlo Tree Diffusion — System 2 Planning with Diffusion Models
Paper review of Monte Carlo Tree Diffusion (Yoon, Cho et al., ICML 2025 Spotlight) — a framework that embeds MCTS inside the diffusion denoising process, treating partially denoised subplans as tree nodes, binary guidance schedules as meta-actions, and fast DDIM rollouts as the simulation oracle. Near-perfect success on long-horizon maze tasks where all diffusion baselines fail, with monotonically improving performance as inference compute scales.
-
Video Prediction Policy — Predictive Visual Representations as the Policy Backbone
Paper review of Video Prediction Policy (Hu, Guo et al., ICML 2025 Spotlight) — a generalist robot policy that conditions on the internal predictive representations of a fine-tuned video diffusion model and learns an implicit inverse-dynamics action head on top. +41.5% relative improvement on Calvin ABC-D and a single policy solving 100+ real-world dexterous-hand tasks.
-
DreamGen — Scaling Robot Learning by Dreaming Trajectories
Paper review of DreamGen (Jang, Ye et al., NVIDIA GEAR, CoRL 2025) — a four-stage pipeline that fine-tunes a video world model on a small amount of teleoperation data, dreams synthetic robot videos for new behaviors and environments, recovers pseudo-actions via IDM or latent-action models, and trains visuomotor policies on the resulting "neural trajectories." Enables a humanoid to perform 22 new behaviors from just a single pick-and-place teleoperation source task.