Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning

TL;DR

Proposes Token Assorted, mixing latent and text tokens to improve reasoning, reducing trace length by 17% and boosting performance by up to 13.3%.

cs.CL 🔴 Advanced 2025-02-05 38 views
DiJia Su Hanlin Zhu Yingchen Xu Jiantao Jiao Yuandong Tian Qinqing Zheng
language models reasoning optimization latent tokens hybrid representation mathematical reasoning

Key Findings

Methodology

Introduces a hybrid representation where initial reasoning steps are abstracted into discrete latent tokens via VQ-VAE. A randomized mixing strategy enables fast adaptation to new latent tokens.

Key Results

  • In mathematical reasoning, Llama-3.1-8B achieves a 13.3% improvement on Fresh-Gaokao-Math-2023 with a 22% reduction in trace length.
  • In Keys-Finding Maze, latent token models outperform CoT baselines with a 19.8% accuracy boost.
  • On ProsQA, accuracy improves from 77.5% (CoT baseline) to 96.2% using latent tokens.

Significance

This work significantly reduces reasoning trace lengths while improving performance on logical and mathematical reasoning tasks. It offers a novel approach to efficient reasoning for resource-constrained scenarios.

Technical Contribution

First to integrate discrete latent tokens into reasoning tasks, introducing a randomized mixing strategy for training. Demonstrates effectiveness across multiple benchmarks.

Novelty

Combines discrete latent tokens with text tokens to address CoT inefficiencies. Outperforms continuous latent space methods in both efficiency and accuracy.

Limitations

  • Relies on VQ-VAE quality for latent token generation, potentially affecting accuracy.
  • Generalization to larger real-world datasets remains untested.
  • Compression rate tuning may require task-specific adjustments.

Future Work

Future work includes improving latent token generation, optimizing compression rates, and validating on broader real-world tasks.

AI Executive Summary

Chain-of-thought (CoT) methods have proven effective for reasoning in large language models (LLMs), but their lengthy input sequences lead to high computational costs. To address this, the authors propose a hybrid representation that abstracts initial reasoning steps into discrete latent tokens generated by VQ-VAE, significantly reducing trace lengths.

The method employs a randomized mixing strategy during training, enabling models to adapt quickly to new latent tokens. Experimental results highlight its effectiveness: a 13.3% performance boost on Fresh-Gaokao-Math-2023, a 19.8% accuracy improvement on Keys-Finding Maze, and a leap from 77.5% to 96.2% accuracy on ProsQA.

This approach not only enhances reasoning efficiency but also opens new avenues for resource-efficient LLM applications. However, challenges remain in latent token quality and generalization to larger datasets. Future research will focus on optimizing token generation and exploring broader applications.

Deep Analysis

Background

CoT methods improve reasoning by breaking problems into explicit steps but suffer from high computational costs due to long input sequences. Recent efforts to compress reasoning traces using latent spaces have shown promise but often require multi-stage training, which remains resource-intensive.

Core Problem

The core challenge is reducing reasoning trace lengths without sacrificing performance. Existing CoT methods are computationally expensive, and latent space approaches lack efficiency and generalization.

Innovation

This paper introduces a hybrid representation combining discrete latent tokens and text tokens. Latent tokens abstract initial reasoning steps, reducing trace lengths. A randomized mixing strategy ensures models adapt effectively to new tokens.

Methodology

  • �� Use VQ-VAE to generate discrete latent tokens, abstracting initial reasoning steps.
  • �� Randomly mix latent and text tokens during training to enhance adaptability.
  • �� Directly generate latent tokens during inference, eliminating VQ-VAE overhead.

Experiments

Experiments span logical reasoning (Keys-Finding Maze, ProntoQA, ProsQA) and mathematical reasoning (Math, GSM8K). Metrics include accuracy and trace length reduction. Ablation studies validate the method's components.

Results

On Keys-Finding Maze, accuracy improves by 19.8%. On ProsQA, accuracy rises from 77.5% to 96.2%. On Fresh-Gaokao-Math-2023, performance improves by 13.3%.

Applications

Applicable to resource-efficient reasoning tasks like mathematical problem-solving, logical inference, and multi-step planning.

Limitations & Outlook

Performance depends on VQ-VAE quality. Generalization to larger datasets is untested. Compression rate tuning may require task-specific optimization.

Plain Language Accessible to non-experts

Imagine you're assembling furniture. Traditional methods involve detailed step-by-step instructions, like CoT. This paper's approach simplifies it into high-level steps, like 'assemble frame' or 'attach legs,' saving time while still achieving the same result.

ELI14 Explained like you're 14

Think of solving a puzzle game where you need keys to unlock doors. Normally, you'd write down every step, like 'go to Room A, pick up the red key.' This method skips the details and just says 'got the key,' so you finish faster! Cool, right?

Glossary

VQ-VAE

A generative model that compresses input into discrete latent representations.

Used to generate latent tokens for reasoning trace compression.

CoT (Chain-of-Thought)

A reasoning method that breaks problems into explicit steps.

Baseline method compared to the proposed approach.

Latent Token

Discrete representations generated by VQ-VAE to replace text tokens.

Used to reduce reasoning trace lengths.

Keys-Finding Maze

A reasoning task requiring planning to find keys and unlock doors.

Evaluates planning capabilities.

GSM8K

A dataset of complex mathematical problems.

Used to evaluate mathematical reasoning performance.

Open Questions Unanswered questions from this research

  • 1 How to further improve latent token quality?
  • 2 Can the method generalize to larger, real-world datasets?
  • 3 Is dynamic compression rate adjustment feasible?

Applications

Immediate Applications

Mathematical Problem-Solving

Optimizes computational efficiency in math reasoning tasks, useful in education.

Logical Inference

Enhances performance in logical reasoning tasks, applicable to QA systems.

Long-term Vision

General Reasoning Optimization

Extends efficient reasoning methods to domains like healthcare and scientific research.

Abstract

Large Language Models (LLMs) excel at reasoning and planning when trained on chainof-thought (CoT) data, where the step-by-step thought process is explicitly outlined by text tokens. However, this results in lengthy inputs where many words support textual coherence rather than core reasoning information, and processing these inputs consumes substantial computation resources. In this work, we propose a hybrid representation of the reasoning process, where we partially abstract away the initial reasoning steps using latent discrete tokens generated by VQ-VAE, significantly reducing the length of reasoning traces. We explore the use of latent trace abstractions in two scenarios: 1) training the model from scratch for the Keys-Finding Maze problem, 2) fine-tuning LLMs on this hybrid data with an extended vocabulary including unseen latent tokens, for both logical and mathematical reasoning problems. To facilitate effective learning, we introduce a simple training procedure that randomly mixes latent and text tokens, which enables fast adaptation to new latent tokens. Our approach consistently outperforms the baselines methods in various benchmarks.

cs.CL cs.AI cs.LG cs.LO