- llm
- nlp
- reasoning
- rl
- rlhf
- ssm
- theory
- information-theory
- representation-learning
- optimization
- rag
- alignment
- paper-review
- survey
- notes
•
•
•
•
•
•
•
•
•
•
•
•
•
•
-
Huginn — Scaling Test-Time Compute via Recurrent Depth in Latent Space
Paper review of "Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach" (Geiping, McLeish, Jain, Kirchenbauer, Singh, Bartoldson, Kailkhura, Bhatele, Goldstein — ELLIS Institute Tubingen, University of Maryland, Lawrence Livermore National Laboratory) — the "Huginn" model, which scales test-time compute by looping a shared recurrent block in latent space instead of emitting explicit chain-of-thought tokens, trained at 3.5B parameters / 800B tokens on the Frontier supercomputer.
-
LoopFormer — Elastic-Depth Looped Transformers via Shortcut Modulation
Paper review of LoopFormer (Jeddi, Ciccone, Taati, ICLR 2026, University of Toronto / Vector Institute / UHN) — a looped transformer trained on variable-length recurrence trajectories so a single set of weights supports elastic inference-time depth. A shortcut-consistency training scheme distills coarse-schedule representations toward fine-schedule ones, letting users trade compute for quality at inference without retraining.
-
Think-at-Hard — Selective Latent Iteration for Looped Reasoning Transformers
Paper review of Think-at-Hard / TaH (Fu, You, Chen, Dai, Yang, Wang, Tsinghua University) — a looped transformer that learns to run a second latent iteration only on tokens likely to be wrong after the first pass, via a lightweight decider, duo-causal attention across the (position, depth) grid, and depth-aware LoRA adapters. TaH gains 8-11% accuracy over a fixed-two-iteration baseline while skipping 94% of second passes.
-
Process Reward Agents — Online Step-Wise Steering for Knowledge-Intensive Reasoning
Paper review of PRA (Sohn, Sternal et al., ICML 2026, ETH Zürich / Heidelberg) — a framework that trains a separate reward agent to provide online, process-level, retrieval-grounded feedback to a frozen policy model during beam search. A 0.5B–8B frozen reasoner gets up to +25.7% on MedQA; Qwen3-4B-Instruct reaches 80.8% SOTA at 4B scale. Self-Consistency degrades on hard benchmarks; PRA continues improving as inference compute scales.
-
Tele-Lens — How Far Ahead Do LLMs Actually Plan in Chain-of-Thought?
Paper review of Tele-Lens (Xu, Yu et al., ICML 2026, WeChat AI / Tencent) — a probing framework that measures how far ahead LLMs plan in hidden states during chain-of-thought reasoning. The answer: myopic. Final-answer probability in hidden states stays near chance (0.49–0.51) throughout CoT for compositional tasks, spiking to 0.94–0.99 only in the last 1–2 steps. Leveraging this, pivot-based uncertainty estimation improves accuracy +6%, and automatic CoT bypass achieves 16.2% bypass rate with negligible performance loss.