7 papers

Reasoning & Test-Time Compute

Chains, trees, atoms, and System-2 thinking that trade compute for correctness.

Papers
7
Read time
2.0h
Avg length
17m
01
Reasoning & Test-Time Compute

Hierarchical Reasoning Model

A 27M-parameter brain-inspired recurrent network that reasons inside its hidden state — no chain-of-thought, no pretraining, just 1000 examples — and beats giant LLMs on Sudoku, mazes, and ARC-AGI.

GUAN WANG, JIN LI, YUHAO SUN, SEN SONG, YASIN ABBASI YADKORI ET AL. (SAPIENT INTELLIGENCE) · 2025
Read
22 min
Source
arXiv:2506.21734
02
Reasoning & Test-Time Compute

Atom of Thoughts for Markov LLM Test-Time Scaling

Instead of dragging the entire reasoning history forward at every step, AOT repeatedly rewrites a hard question into a smaller, self-contained "atomic" question — so the model spends all its compute on the problem in front of it, not on re-reading its own past work.

TENG, YU, SHI, ZHANG, WU, LUO (HKUST-GZ · DEEPWISDOM · RENMIN UNIVERSITY) · FEB 2025
Read
16 min
Source
arXiv:2502.12018
03
Reasoning & Test-Time Compute

Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models

Insight-V teaches vision-language models to reason like OpenAI o1 by (a) auto-generating long step-by-step reasoning data with no human labelers and (b) splitting the job across two specialized agents — one that reasons and one that judges-and-summarizes — instead of asking a single model to do both.

YUHAO DONG, ZUYAN LIU, HAI-LONG SUN, ET AL. (NTU S-LAB · TENCENT · TSINGHUA · NANJING) · CVPR 2025
Read
16 min
Source
arXiv:2408.12637
04
Reasoning & Test-Time Compute

Proof of Thought: Neurosymbolic Program Synthesis for Robust and Interpretable Reasoning

Instead of trusting an LLM's free-text reasoning, make it emit a structured JSON "program" of facts and rules, compile that into formal logic, and let a theorem prover (Z3) decide the answer — so the conclusion is *provable*, not just plausible.

DEBARGHA GANGULY, SRINIVASAN IYENGAR, VIPIN CHAUDHARY, SHIVKUMAR KALYANARAMAN (CASE WESTERN / MICROSOFT) · NEURIPS 2024 WORKSHOP
Read
16 min
Source
arXiv:2409.17270
05
Reasoning & Test-Time Compute

Reasoning Language Models for Root Cause Analysis in 5G Wireless Networks

Take a small open-weight LLM, teach it to diagnose 5G network faults by first feeding it structured reasoning traces (SFT) and then rewarding it for correct diagnoses (GRPO), and a 1.5B model beats a 70B reasoning model at telecom root-cause analysis.

SANA, PIOVESAN, DE DOMENICO, KANG, ZHANG, DEBBAH, AYED (HUAWEI / KHALIFA UNIVERSITY) · ARXIV 2025 · 2507.21974V1
Read
16 min
06
Reasoning & Test-Time Compute

The Optimal Choice of Hypothesis Is the Weakest, Not the Shortest

When you infer a rule from a few examples, the rule most likely to keep working on new cases is not the *shortest* one (Occam's Razor / compression) but the *weakest* one — the rule that commits to the least and stays compatible with the most possibilities.

MICHAEL TIMOTHY BENNETT
Read
14 min
Source
arXiv:2301.12987
07
Reasoning & Test-Time Compute

Towards System 2 Reasoning in LLMs: Learning How to Think With Meta Chain-of-Thought

The "thinking out loud" that o1/R1-style models do at inference time is best understood as an *internalized search* over solutions — and this paper lays out the theory plus a concrete train-it-yourself pipeline (synthetic search traces + reinforcement learning) for getting a single model to do that search by itself.

XIANG, SNELL, GANDHI, RAFAILOV, FINN ET AL. (SYNTHLABS · STANFORD · UC BERKELEY)
Read
22 min
Source
arXiv:2501.04682