PaperlayerView exact revision on arXiv

Source-linked AI summary

ReGuLaR: Variational Latent Reasoning Guided by Rendered Chain-of-Thought

Fanmeng Wang, Haotian Liu, Guojiang Zhao, Hongteng Xu, Zhifeng Gao

arXiv:2601.23184v1cs.CL

TL;DR

Explicit CoT improves reasoning but is computationally redundant, while existing latent methods can lose performance without suitable compression guidance. ReGuLaR combines VAE-based latent reasoning with rendered CoT visual representations to regularize latent states. It outperforms existing latent methods in efficiency and reasoning effectiveness and can surpass explicit CoT through multi-modal reasoning.

  • Problem

    Explicit CoT introduces computational redundancy, while existing latent reasoning methods can suffer severe performance degradation from inadequate compression guidance.

  • Method

    ReGuLaR formulates latent reasoning within a VAE framework and uses visual representations extracted from rendered explicit reasoning chains to regularize posterior latent states.

  • Results

    ReGuLaR significantly outperforms existing latent reasoning methods across computational efficiency and reasoning effectiveness, and surpasses explicit CoT through multi-modal latent reasoning.

  • Takeaways & Limitations

    Rendered CoT provides a semantically meaningful, information-preserving prior for compact latent reasoning states.

  • Takeaways & Limitations

    Existing latent reasoning methods remain limited by the lack of appropriate compression guidance, which can cause severe performance degradation compared with explicit CoT.

Abstract

from arXiv · show

While Chain-of-Thought (CoT) significantly enhances the performance of Large Language Models (LLMs), explicit reasoning chains introduce substantial computational redundancy. Recent latent reasoning methods attempt to mitigate this by compressing reasoning processes into latent space, but often suffer from severe performance degradation due to the lack of appropriate compression guidance. In this study, we propose Rendered CoT-Guided variational Latent Reasoning (ReGuLaR), a simple yet novel latent learning paradigm resolving this issue. Fundamentally, we formulate latent reasoning within the Variational Auto-Encoding (VAE) framework, sampling the current latent reasoning state from the posterior distribution conditioned on previous ones. Specifically, when learning this variational latent reasoning model, we render explicit reasoning chains as images, from which we extract dense visual-semantic representations to regularize the posterior distribution, thereby achieving efficient compression with minimal information loss. Extensive experiments demonstrate that ReGuLaR significantly outperforms existing latent reasoning methods across both computational efficiency and reasoning effectiveness, and even surpasses CoT through multi-modal reasoning, providing a new and insightful solution to latent reasoning. Code: https://github.com/FanmengWang/ReGuLaR.

1 Introduction

CoT improves difficult-problem solving but incurs token-level redundancy and latency. ReGuLaR addresses latent reasoning’s information-loss challenge by using a VAE framework and rendered visual representations to regularize latent states, achieving efficient and effective reasoning.

  • Motivation: Explicit CoT improves accuracy on challenging queries but creates computational overhead and inference latency through redundant token-by-token generation.The redundancy limits the scalability of LLM reasoning.
  • Motivation: Latent reasoning reduces decoding overhead by compressing reasoning into continuous representations, but unconstrained recursion can accumulate errors and cause semantic drift.Existing methods therefore face an information-preservation challenge despite their computational benefits.
  • ReGuLaR: ReGuLaR formulates latent reasoning as probabilistic modeling within a VAE framework, sampling each latent state from a posterior conditioned on previous states.The model optimizes an ELBO, with the prior regularizing the posterior distribution.
  • ReGuLaR: Rendered explicit reasoning chains provide dense visual-semantic representations that regularize the latent-state posterior during training.The reasoning chain is rendered as images and encoded visually for this regularization.
  • Reported outcome: ReGuLaR achieves state-of-the-art performance with minimal reasoning length and supports multi-modal latent reasoning that can surpass explicit CoT in complicated scenarios.The paper presents this as a new solution to latent reasoning.

2 Related Work

Related work reduces reasoning or context costs through explicit CoT variants, latent representations, and visual-text compression. ReGuLaR extends these directions by rendering reasoning into images, using dense visual representations as compression guidance.

  • Explicit reasoning: CoT variants expose intermediate reasoning in different explicit forms, but verbose steps incur substantial computational cost and inference latency.The related approaches include Tab-CoT, ToT, and GoT-Rationale.
  • Latent reasoning: Latent reasoning replaces intermediate reasoning tokens with continuous representations, eliminating language-decoding overhead during latent reasoning.Coconut pioneers this direction by recursively using hidden states as continuous latent thoughts.
  • Latent reasoning: Existing latent methods can suffer severe performance degradation because they lack appropriate compression guidance.This limitation restricts their practical utility relative to explicit CoT.
  • Visual-text compression: Visual-text compression renders textual content as images and embeds it with a visual encoder to exploit the visual modality’s information density.Prior work mainly compresses input contexts rather than reasoning processes.
  • Visual-text compression: Visual representations are presented as high-density carriers of textual information, motivating their use for empowering latent reasoning.ReGuLaR applies this visual-text compression idea to latent reasoning.

3 Proposed Method

The method replaces explicit chain-of-thought generation with shorter probabilistic latent states and trains them within a variational auto-encoding framework. ReGuLaR regularizes these states using visual representations of rendered reasoning segments, preserving semantic information without adding inference-time cost.

  • Problem Statement and Preliminaries: Explicit CoT generates intermediate reasoning tokens token by token, creating substantial computational overhead and inference latency.Latent reasoning instead replaces the discrete reasoning chain with continuous states that can be much shorter than the original chain.
  • The Variational Latent Reasoning Framework: ReGuLaR models latent reasoning states probabilistically, sampling each state from a posterior conditioned on the question and preceding latent states.The latent reasoning head produces distribution parameters and samples states through reparametrization.
  • The Variational Latent Reasoning Framework: The VAE objective combines reasoning-token likelihood with KL divergence that regularizes the posterior toward a prior distribution.The autoregressive module with the latent reasoning head acts as an encoder, while the language head decodes tokens from latent states.
  • Implementing the Framework via ReGuLaR: ReGuLaR renders each reasoning segment as an image, extracts visual representations with a pretrained encoder, and maps them into the latent reasoning space through a trainable adapter.These visual representations parameterize the prior distribution used to regularize latent reasoning states.
  • Implementing the Framework via ReGuLaR: Rendering preserves semantic integrity better than grouped token-embedding aggregation, while inference accepts pure text inputs and adds no extra computational cost.The framework can also encode charts, graphs, and diagrams alongside text for multimodal latent reasoning.

4 Experiments

Experiments evaluate ReGuLaR across math datasets, backbones, compression rates, and model scales. ReGuLaR achieves stronger accuracy and shorter reasoning than baselines, including under extreme compression.

  • Experimental Setup: Experiments use GSM8K-Aug, three out-of-domain math datasets, and GSM8K-Aug-NL, with Accuracy and Reasoning Length evaluated over five random seeds.Baselines include iCoT, CODI, Coconut, and CoLaR under a unified framework.
  • Performance Comparison: ReGuLaR achieves state-of-the-art performance across four math reasoning datasets, improving accuracy while reducing average reasoning length by approximately 35%, from 4.70 to 3.03.The comparison uses CoLaR as the strongest baseline.
  • Generalizability Analysis: With DeepSeek-R1-Distill-Qwen-1.5B, ReGuLaR maintains the highest accuracy and shortest reasoning length across datasets.On GSM-Hard, it achieves higher accuracy than CoLaR with 3.1 reasoning steps instead of 12.8.
  • Compression Analysis: ReGuLaR consistently outperforms CoLaR across compression rates on both LLaMA-3.2-1B-Instruct and DeepSeek-R1-Distill-Qwen-1.5B backbones.The compression rate is the number of explicit reasoning tokens represented by one latent reasoning state.
  • Scalability Analysis: ReGuLaR maintains a significant performance margin over CoLaR and Coconut across LLaMA-3 model scales from 1B to 8B and across datasets.The evaluated backbones are LLaMA-3.2 1B/3B and LLaMA-3.1 8B Instruct variants.
  • Extreme Compression: Under single-step extreme compression, ReGuLaR outperforms CoLaR across all model scales and datasets.On MATH, average accuracy rises from 7.76% to 11.9% while reasoning length falls from 62.2 to 1.00.

5 Latent Reasoning Beyond Textual Domain

ReGuLaR extends latent reasoning to multimodal intermediate chains while retaining textual input and output. On molecular captioning, it achieves state-of-the-art results with one latent reasoning step and surpasses both CoLaR and explicit CoT.

  • Dataset: The molecular captioning benchmark uses SELFIES strings as inputs and requires natural-language molecule descriptions as outputs.The benchmark preserves textual I/O while incorporating multimodal reasoning chains as intermediate bridges.
  • Baselines and Evaluation: ReGuLaR renders the entire multimodal reasoning chain into one image, while CoT and CoLaR train on the original textual chains.Performance is evaluated with BLEU, METEOR, and ROUGE.
  • Results: With a single latent reasoning step, ReGuLaR achieves state-of-the-art performance across all metrics and backbones.It significantly outperforms both CoLaR and explicit CoT, which use hundreds of reasoning steps.

6 Conclusion and Future Work

ReGuLaR presents a VAE-based latent reasoning paradigm guided by rendered CoT and reports improved efficiency and reasoning ability. The paper evaluates this approach across mathematical, algebraic, and multimodal reasoning settings while identifying limitations in current benchmarks.

  • ReGuLaR models latent reasoning within the VAE framework and learns it using rendered CoT guidance.
  • ReGuLaR significantly outperforms existing latent reasoning methods in computational efficiency and reasoning ability, and surpasses explicit CoT through multimodal latent reasoning.
  • Future Work: Current benchmarks such as GSM8K and GSM8K-Aug may inadequately assess advanced reasoning because of limited data sizes and overly simple reasoning chains.
  • Future Work: Future work will develop larger, higher-quality reasoning datasets and investigate whether and how latent reasoning can outperform explicit CoT in theory.
  • The method is evaluated on GSM8K-Aug, GSM8K-Aug-NL, GSM-Hard, SVAMP, MultiArith, AQUA-RAT, and MATH.
  • The experiments compare ReGuLaR with iCoT, CODI, Coconut, and CoLaR as latent-reasoning baselines.

A.3 Implementation

ReGuLaR renders reasoning segments, encodes them as visual-semantic representations, and trains a VAE-based latent reasoning model. Its inference procedure samples latent states iteratively before generating the final answer.

  • Rendering Configuration: The rendering function maps token sequences to images using a configuration selected for consistent visual encoding and high semantic density.
  • Visual Encoder: A visual encoder converts rendered images into continuous visual-semantic representations used in the latent reasoning training pipeline.
  • Training Scheme: Training first divides each reasoning chain into segments, renders each segment as an image, extracts visual representations, and stores them for training.
  • Training Scheme: Training combines latent reasoning loss, KL regularization, and answer-generation loss before updating the trainable parameters.
  • Hyperparameters: The implementation primarily uses a frozen LLaMA-3.2-1B-Instruct backbone with trainable LoRA modules, adapter, and latent reasoning head.
  • Inference Scheme: Inference predicts posterior parameters from the question and latent-state history, samples a latent state, decodes a representative token, and stops at an end-of-reasoning token.

B.1 Ablation Studies on Rendering Configuration

ReGuLaR remains stable across broad font-size and rendering-density settings. The reported ablations indicate that performance varies only modestly without precise rendering-parameter tuning.

  • Font Size: 44.5% to 45.6% average accuracy is observed across font sizes from 9pt to 20pt, while reasoning length remains largely consistent.
  • Rendering Density: 45.6% average accuracy at 72 DPI and 45.2% at 300 DPI demonstrate comparable performance across the tested density range.
  • Rendering Density: The density ablation varies DPI from 72 to 300 and reports averaged Accuracy and Reasoning Length with 95% confidence intervals.
  • Font Size: Font-size ablations report averaged Accuracy and Reasoning Length with 95% confidence intervals.

B.2 Ablation Studies on Visual Encoder Modes

The visual encoder supports four resolution modes whose token sequences are mean-pooled into compact representations. The Tiny mode performs comparably to higher-resolution modes despite using fewer intermediate tokens.

  • Visual Encoder Modes: Tiny, Small, Base, and Large modes use 512 × 512, 640 × 640, 1024 × 1024, and 1280 × 1280 resolutions, respectively.
  • Visual Encoder Modes: The four encoder modes produce 64, 100, 256, and 400 visual tokens before mean pooling into a single visual-semantic representation.
  • Visual Encoder Modes: Tiny mode achieves performance comparable to high-resolution modes despite resizing inputs to 512 × 512 and using only 64 intermediate tokens.
  • Visual Encoder Modes: The reported robustness is attributed to information aggregation through pooling of intermediate visual tokens.

B.3 Ablation Studies on Learning Paradigms

The ablations show that posterior regularization is essential: the KL term prevents catastrophic failure, while combining semantic and distributional supervision achieves the best performance.

  • Learning objective: The unified objective combines answer generation, latent reasoning, and KL-divergence losses, with answer generation retained across ablation variants.These components respectively target answer correctness, semantic integrity, and posterior regularization.
  • Ablation results: Accuracy< 14% occurs without the KL divergence term, regardless of whether the latent reasoning loss is included.The result indicates that removing posterior regularization causes catastrophic failure in this ablation.
  • Ablation results: KL alone raises performance to 41.9%, showing that distribution regularization provides a critical constraint for latent reasoning.The paper contrasts this constraint with supervision from the final answer and textual reconstruction.
  • Ablation results: Combining all components reaches 45.6%, indicating complementary contributions from text reconstruction and distribution regularization.The paper describes their semantic richness and structural guidance as synergistic and mutually indispensable.

B.4 Ablation Studies on Modeling Strategies

ReGuLaR’s probabilistic latent-state transitions outperform deterministic prediction across datasets. The authors attribute this advantage to avoiding mean collapse and preserving distinct semantic reasoning states.

  • Modeling strategies: The probabilistic strategy predicts distribution parameters µ and log σ for the next latent reasoning state, whereas the deterministic variant predicts that state directly.The deterministic version is functionally equivalent to greedily selecting the most likely mean vector.
  • Modeling strategies: 45.6% average accuracy is achieved by probabilistic modeling, which outperforms the deterministic variant across all datasets.The comparison is reported in Table 9.
  • Interpretation: Deterministic prediction can cause mean collapse by averaging multiple valid subsequent reasoning steps into blurred semantic representations.This averaging minimizes reconstruction error but fails to capture precise logic for complex reasoning.
  • Interpretation: Probabilistic modeling samples sharp, distinct latent reasoning states from the underlying distribution, preserving semantic integrity.This is the authors’ explanation for its stronger performance than deterministic prediction.

B.5 Ablation Studies on Regularization Strategies

Vision-based regularization outperforms text-based regularization by providing compact, structured semantic anchors for latent reasoning. The approach also maintains strong scaling behavior across model sizes and datasets.

  • Regularization strategies: Rendered reasoning segments produce visual representations that regularize the posterior distribution as dense semantic anchors during training.The text-based alternative aggregates token embeddings within each segment while keeping other settings invariant.
  • Regularization strategies: The vision-based strategy raises average accuracy from 42.3% to 45.6% compared with text-based regularization across all datasets.Text pooling can dilute structural and topological details, whereas rendered images preserve a more holistic segment representation.
  • Efficiency: The visual representations are confined to training, so vision-based regularization adds no inference cost.The method aligns latent states with compact rendered images during optimization rather than decoding them at inference.
  • Scalability: ReGuLaR maintains the highest accuracy and shortest reasoning length across model scales and datasets compared with CoLaR and Coconut.Figure 6 reports this positive scaling behavior using LLaMA-3.2 1B and 3B and LLaMA-3.1 8B Instruct backbones.
  • Multimodal reasoning: In molecular captioning, rendered multimodal reasoning chains pair textual reasoning with explicit 2D molecular graphs before visual extraction.RDKit generates the corresponding molecular graphs, and Figure 7 shows rendered and original chain variants.
Loading 2601.23184v1…