- llm
- nlp
- reasoning
- rl
- rlhf
- ssm
- theory
- information-theory
- representation-learning
- optimization
- rag
- alignment
- paper-review
- survey
- notes
•
•
•
•
•
•
•
•
•
•
•
•
•
•
-
The Expressive Power of Transformers with Chain of Thought
Paper review of Merrill & Sabharwal (ICLR 2024, NYU / Allen Institute for AI) — a circuit-complexity characterization of what decoder-only transformers can compute once they are allowed to generate intermediate chain-of-thought tokens, and exactly how many of those tokens it takes to buy how much extra power.
-
Landscape of Thoughts — Visualizing Where LLM Reasoning Actually Goes
Paper review of Landscape of Thoughts (Zhou, Zhu, Li, Galkin, Feng, Koyejo, Tang, Han — ICLR 2026, TMLR Group @ HKBU / Stanford / Mila / Université de Montréal / HEC Montréal / Intel AI Lab) — the first visualization tool that projects every intermediate reasoning state of an LLM's chain-of-thought trajectory into a 2D "landscape" relative to the answer choices, turning perplexity-based distance into a picture of how (and whether) reasoning converges toward the right answer.
-
Magellan — Guided MCTS for Escaping the Gravity Wells of LLM Creativity
Paper review of Magellan (Lufan Chang, agents4science 2025) — a framework that reframes creative idea generation as guided MCTS over an LLM's latent conceptual space, replacing unreliable self-evaluation heuristics (as used in Tree-of-Thoughts) with a hierarchical guidance system: a global "semantic compass" and a local landscape-aware value function combining coherence, novelty, and progress.
-
PriorZero — Injecting LLM Priors into MuZero-Style World Models at the MCTS Root
Paper review of PriorZero (Xiong, Pu, Tang, Niu) — a method for combining LLM semantic priors with UniZero-style latent world models by blending the LLM's policy prior into MCTS only at the root node, preserving the world model's own deep lookahead while cheaply steering exploration toward semantically plausible actions. Strong gains on sparse-reward Jericho text-adventures and compositional generalization in BabyAI.
-
SuperThoughts — Reasoning Tokens in Superposition
Paper review of SuperThoughts (Xiong, Garg, Yu, Shrivastava, Zhao, Kyrillidis, Papailiopoulos, ICML 2026, UW-Madison / Microsoft Research / Princeton / Rice) — compressing pairs of consecutive chain-of-thought tokens into a single latent representation and decoding two tokens per step via a lightweight multi-token-prediction module, with a confidence-gated fallback to standard decoding on hard reasoning steps. ~20-30% CoT length reduction near accuracy parity.