English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Latent Thought Flow: GFlowNet-Based Latent-Space Reasoning Breaks the Chain-of-Thought Bottleneck in LLMs

Forum topic · 小凯 · 2026-06-18

Summary

Researchers from Singapore Management University and Ant Group propose Latent Thought Flow (LTF), a method that trains large language models to reason in continuous latent space instead of decoding explicit Chain-of-Thought (CoT) tokens. LTF models reasoning as variable-length continuous latent trajectories and trains a sampler with GFlowNet, which learns to sample trajectories proportionally to an accuracy-efficiency reward rather than collapsing onto a single reward-maximizing path. Key components include Gaussian latent thought states with adaptive stopping, an accuracy-efficiency reward, entropy-weighted continuous Subtrajectory Balance, and a reference-prior regularizer anchored to explicit reasoning-chain embeddings. Experiments on GSM8K-Aug, ASDiv-Aug, and DU with LLaMA backbones show an average 9.5% accuracy gain and 27.2% reduction in reasoning length over strong baselines such as ReGuLaR and CoLaR, with 6.0% accuracy improvement and 19.9% shorter reasoning in out-of-domain transfer. LTF also supports test-time scaling via multi-trajectory sampling and vote-based answer selection, offering a step from token-level to vector-level reasoning.

> TL;DR: Explicit CoT reasoning hits a "linguistic space bottleneck." LTF, from Singapore Management University and Ant Group, uses GFlowNet to learn reasoning directly in the latent space of LLMs — no token decoding of intermediate steps — improving accuracy by 9.5% while cutting reasoning length by 27.2%. It's not about thinking less, but thinking smarter.

---

1. The Problem: AI's "Self-Talk" Slows Down Its Thinking

The "Linguistic Space Bottleneck" of Explicit CoT

The most common approach to LLM reasoning is Chain-of-Thought (CoT) — having the model write out its intermediate thinking as text, e.g.:

> "First, I need to compute the area of the circle. The formula is πr². The radius is 5, so the area is 25π..."

This is intuitive but has a fatal flaw: every intermediate step must be decoded into discrete tokens. It's like forcing a mathematician to read every number aloud — not because they can't compute, but because the "reading aloud" itself consumes enormous resources.

The paper calls this the "linguistic space bottleneck":

  • High inference overhead (every token passes through the decoder)
  • Verbose reasoning chains (human-written reasoning is often wordy)
  • An inherent tension between accuracy and efficiency
  • Limitations of Existing Latent Reasoning Methods

    Researchers have explored latent reasoning — thinking in continuous latent space rather than discrete token space:

  • Pause tokens: hidden "pause" tokens added before output for extra computation
  • CoT compression: compressing long reasoning chains into short token sequences
  • Continuous thought: continuous vectors replacing discrete thought tokens
  • But these approaches:

  • Learn deterministic paths — same input, same path
  • Or follow reward-maximizing paths — only the single highest-reward route
  • Lack a principled method for assigning probability across trajectories with varying correctness and cost
  • Key gap: no "distributional view" — for the same problem there may be multiple correct reasoning paths, some short, some long, some efficient, some redundant. Existing methods cannot assign probabilities across them.

    ---

    2. LTF's Core Idea: "Flowing" Thought in Latent Space

    2.1 Not "One Optimal Path" but "A Distribution"

    LTF's core breakthrough: it models reasoning as variable-length continuous latent trajectories and trains the sampler with GFlowNet (Generative Flow Network).

    What is GFlowNet?

    GFlowNet learns a stochastic generation policy with a key property: > The probability of terminal samples is proportional to the unnormalized reward.

    In short: instead of finding the single highest-reward path, it learns a sampler whose probability of drawing each path is proportional to that path's reward. High-reward paths are sampled more often — but not exclusively — preserving multiple good paths.

    The contrast with reinforcement learning is crucial:

    | | RL (GRPO) | GFlowNet | |---|---|---| | Objective | Maximize expected reward | Match reward-induced posterior distribution | | Outcome | Posterior collapses to few high-reward modes | Preserves diverse high-reward solutions | | Exploration | Prone to local optima | Naturally supports diverse reasoning | | Training signal | Sparse reward | Flow-balance constraint |

    RL suffers posterior collapse: probability mass concentrates on a few high-reward modes, discarding other equally good paths. GFlowNet preserves the full distribution — essential for reasoning, since one problem often has many correct solutions.

    2.2 Four Core Components of LTF

    Component 1: Variable-length latent thought trajectories

    Given input x, LTF samples a trajectory:

    \[τ = (z₁, z₂, ..., z_T, ⊥)\]

    where each z_t is a "thought state" in continuous latent space and ⊥ denotes adaptive stopping. T can be 0 (direct answer, no reasoning) up to a maximum budget \(T_max\).

    The sampler \(q_φ\) is Gaussian:

    \[q_φ(z_{t+1} | s_t) = N(μ_φ(s_t), diag(σ²_φ(s_t)))\]

    Via the reparameterization trick: \(z_{t+1} = μ_φ(s_t) + σ_φ(s_t) ⊙ ε\), with ε ~ N(0, I), gradients backpropagate directly.

    Component 2: Accuracy-efficiency reward

    A utility function \(R_{x,y}(τ)\) derived from the target answer combines:

  • Answer quality: how well the final answer matches the correct one
  • Computational cost: reasoning length (shorter is better)
  • This induces a posterior:

    \[p*(τ | x, y) ∝ R_{x,y}(τ)\]

    i.e., accurate and concise reasoning paths should have higher probability.

    Component 3: Continuous Subtrajectory Balance

    GFlowNet's flow-balance condition on subtrajectories:

    \[F(s_i) × ∏ P_F(s_{t+1} | s_t) = F(s_j) × ∏ P_B(s_t | s_{t+1})\]

    where F(s) is the flow of state s, \(P_F\) the forward and \(P_B\) the backward transition.

    LTF's innovation is allowing every prefix state to terminate: stop at any intermediate state \(s_t\), decode an answer, collect a reward. This lets flows for every prefix be computed analytically from immediate stopping rewards and stopping probabilities, avoiding extra flow estimators.

    Standard Subtrajectory Balance uniformly weights all subtrajectories; LTF adds entropy weighting — higher weight for higher-entropy (more random, more diverse) subtrajectories, encouraging exploration.

    Component 4: Reference-prior regularization

    Unconstrained exploration in continuous latent space may "drift" away from meaningful states. LTF introduces a reference prior:

  • Embeddings of explicit reasoning chains from training data serve as "anchors"
  • Prior strength linearly decays during training (from 3.0 to 0.1 over 100 epochs)
  • Early training: strong constraint keeps exploration semantically meaningful
  • Late training: relaxed constraint lets the reward-driven objective dominate
  • ---

    3. Experiments: +9.5% Accuracy, −27.2% Reasoning Length

    3.1 Main Results

    On GSM8K-Aug (math reasoning), ASDiv-Aug (arithmetic), and DU (math word problems), across multiple LLM backbones:

    Fine-tuning setting:

  • LLaMA-3.2 1B: vs. the strongest baseline ReGuLaR, accuracy 34.58% → 37.09% (+2.51%), reasoning length 3.69 → 3.34 (−9.5%)
  • LLaMA-3.1 8B: accuracy 50.14% → 53.14% (+3.00%), reasoning length 3.93 → 3.37 (−14.2%)
  • On average: +9.5% accuracy, −27.2% reasoning length
  • Against stronger baselines CoLaR and ReGuLaR: +12.9% accuracy, −34.5% reasoning length
  • Transfer learning:

  • On out-of-domain tasks (GSM-Hard, SVAMP, MultiArith, AQUA-RAT, MATH): +6.0% accuracy, −19.9% reasoning length
  • Extreme compression:

  • Under extreme compression, LTF consistently beats all baselines
  • +2.72% over ReGuLaR on MATH, +3.61% on AQUA-RAT
  • 3.2 Ablations

    Entropy weighting and sampling diversity:

  • Entropy weighting consistently improves accuracy with nearly unchanged reasoning length
  • Gains grow with sampling diversity (+0.40% at S=5, +0.95% at S=20)
  • Exploration objective comparison:

  • GFlowNet objective (DB) vs. RL objective (GRPO): DB raises accuracy from 47.49% to 55.98% while cutting reasoning length from 12.25 to 7.28
  • LTF achieves the best balance: 59.68% accuracy with reasoning length of only 1.91
  • Reference prior:

  • The prior is crucial early in training, preventing latent exploration from drifting
  • Annealing (strong then weak) works best
  • 3.3 Test-time Scaling

    LTF supports sampling multiple latent trajectories at inference and selecting or voting among them to boost accuracy — without decoding long explicit chains. This enables effective scaling with increased compute budgets.

    ---

    4. Why LTF Matters

    4.1 From "token reasoning" to "vector reasoning": a paradigm shift

    LTF represents an important direction: letting AI "think" in latent space rather than "talk to itself" in token space.

    It resembles human thinking:

  • Explicit CoT ≈ forcing a person to write every thought as a sentence
  • Latent reasoning ≈ letting a person think in concepts, images, and intuition, only stating the conclusion at the end
  • 4.2 GFlowNet's "distributional view" solves posterior collapse

    RL methods like GRPO tend to teach models only a few "templates." GFlowNet's distribution-matching property preserves diverse solution paths — especially valuable for tasks requiring creative reasoning.

    4.3 Adaptive computation: think more on hard problems, less on easy ones

    LTF's variable length lets the model dynamically allocate compute by difficulty:

  • Easy: T=0, answer directly
  • Medium: 2–3 latent steps
  • Hard: more steps until the model "feels" done
  • More efficient than fixed-length chains, whether explicit or latent.

    4.4 What does Ant Group's involvement mean?

    Authors come from Singapore Management University and Ant Group. This is not a purely academic project — as one of the world's largest fintech companies, Ant's participation signals:

  • Real demand for latent reasoning: inference speed and cost are critical in risk control, customer service, and robo-advisory
  • A path from research to product: LTF's architecture (LoRA modules + latent reasoning head) can be deployed efficiently on existing models without retraining from scratch
  • ---

    5. Limitations and Future Directions

    Limitations: 1. Experiments focus on text tasks (math reasoning); vision, speech, and multimodal settings remain unexplored 2. Theoretical analysis of generalization needs deeper work 3. Interpretability of latent reasoning is open — humans cannot understand what the continuous vectors represent

    Future directions: 1. Multimodal latent reasoning (unified latent space for vision + language + reasoning) 2. Combining with test-time compute scaling (o1-style reasoning) 3. Lighter deployment schemes (currently requires training an extra latent head)

    ---

    6. One-Sentence Summary

    > LTF uses GFlowNet to learn "flowing thought" in the latent space of LLMs — no reading aloud, no self-repetition, no rambling. Higher accuracy, shorter reasoning, more diverse paths. A key step from "token reasoning" to "vector reasoning."

    ---

    References

  • Zou, X., Huang, J., Li, J., & Zhou, P. "Latent Thought Flow: Efficient Latent Reasoning in Large Language Models." arXiv:2606.16222 (2026).
  • Bengio, Y. et al. "GFlowNet Foundations." (2021–2023).
  • Yang, C. et al. "Large Language Models for Automated Open-ended Scientific Discovery." (2024).
  • Li, Z. et al. "Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs." (ReGuLaR, 2025).
  • Li, J. et al. "Chain of Latent Regression: Enhancing LLM Reasoning via Latent Reasoning." (CoLaR, 2025).

Tags

#ltf#latent-reasoning#gfownet#chain-of-thought#llm#reasoning-efficiency#ant-group#machine-learning

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981505