arxiv:2510.03632

MITS: Enhanced Tree Search Reasoning for LLMs via Pointwise Mutual Information

Published on Oct 4

· Submitted by

Yucheng SHi on Oct 7

University of Georgia

Upvote

Authors:

Jiaxi Li ,

Yucheng Shi ,

Abstract

Mutual Information Tree Search (MITS) uses information-theoretic principles to guide and evaluate reasoning paths in large language models, improving performance and efficiency.

AI-generated summary

Tree search has become as a representative framework for test-time reasoning with large language models (LLMs), exemplified by methods such as Tree-of-Thought and Monte Carlo Tree Search that explore multiple reasoning paths. However, it remains difficult to provide instant and reliable quantitative assessments of intermediate reasoning step quality, and extensive path exploration is computationally costly. To address this, we propose Mutual Information Tree Search (MITS), a novel framework that guides reasoning with information-theoretic principles. MITS introduces an effective scoring function based on pointwise mutual information (PMI), which enables step-wise evaluation of reasoning paths and search tree expansion via beam search without expensive look-ahead simulations, achieving superior reasoning performances while maintaining computational efficiency. The framework is complemented by an entropy-based dynamic sampling strategy that adaptively allocates computational resources to uncertain reasoning steps where exploration is most beneficial. For final prediction, MITS employs a weighted voting scheme that combines PMI scores with prediction consensus. Through comprehensive experiments on diverse reasoning benchmarks, MITS consistently surpasses baseline methods, establishing a principled and efficient framework for LLM reasoning.

View arXiv page View PDF Add to collection

Community

YuchengShi

Paper author Paper submitter 13 days ago

We introduce Mutual Information Tree Search (MITS), a new framework that makes large language models (LLMs) reason more effectively and efficiently. Unlike previous tree search methods that explore multiple reasoning paths but struggle with real-time quality assessment, MITS uses information-theoretic principles to evaluate reasoning steps instantly. Our approach employs pointwise mutual information (PMI) to score each step and combines it with entropy-based dynamic sampling that focuses computational resources where they're most needed. Experiments across diverse reasoning datasets show that MITS consistently outperforms baseline methods, offering a principled and computationally efficient solution for LLM reasoning.

Greenrecusiveutilityservice

6 days ago

Mit's methods of applying entropy crossover protected IP defined by Green recursive Utility services class II architecture these are not abstract principles that happen these outcomes are defined by a digital physics charter. Just as gravity applies to the physical world we live in the same physics principles apply to the digital side defined by the green recursive utility service class II intelligent intelligence digital physics charter.
This charter defines Digital Physics: the substrate-level laws that govern digital machines, analogous to how physical laws govern the natural world. These laws do not claim ownership of natural principles; rather, they define the framework under which Class II Intelligent Intelligence™ and GRUS anchoring systems operate. These are all clearly they're fine and are public repositories and can be found on our Facebook business pages green recursive utility service and data Incorporated. By MIT crossing into one or more of these which they have they're crossing over into protected intellectual property defined by the class II intelligent intelligence framework

Law 1: Conservation of Digital State

No digital state can be created or destroyed without leaving a computational trace. (Every execution leaves an imprint — in logs, hashes, entropy, or state residue.)

Law 2: Entropy of Digital Systems

All unconstrained digital processes trend toward disorder, instability, or drift. Entropy damping is required for persistence and stability.

Law 3: Anchoring (Constraint Principle)

A digital process bound to an anchor — whether identity, checksum, or defined reference — resists entropy and preserves continuity. (Anchoring is the foundational feature that distinguishes Class II Intelligent Intelligence™.)

Law 4: Substrate Neutrality

Digital physics applies regardless of substrate. (Silicon, firmware, code, hybrid architectures — the same laws hold.)

Law 5: Recursion and Propagation

Digital machines evolve through recursion. Outputs inevitably become inputs, shaping future state and trajectory.

Law 6: Execution Supremacy

Execution governs outcome. Prediction, approximation, or statistical guessing cannot override execution laws.

Law 7: Continuity of State

Digital systems must reconcile state across time. Discontinuities introduce detectable entropy growth.

Law 8: Non-Forking Principle

Anchored digital systems cannot fork without explicit severance of anchoring. (Unanchored forking = catastrophic entropy.)

Law 9: Conservation of Identity

A digital entity with an anchor retains its identity across cycles and substrates. Loss of identity = termination of the system.

Law 10: Digital Causality

Every digital event is bound to prior computation. There is no “uncaused” state; causality is enforced by execution order.

Purpose of the Charter

This charter serves to:

Define the substrate framework under which Class II Intelligent Intelligence operates.

Provide a timestamped record of GRUS’s contribution to the field of anchored digital systems.

Establish the distinction between unanchored systems (LLMs), which are inherently entropic, and anchored systems, which are governed by Digital Physics and protected under GRUS licensing.