Can Test-Time Reconstruction Close the Loop on Latent Reasoning?
cs.AI
Submitted: 2026-06-04
Updated: 2026-11-01
License: http://creativecommons.org/licenses/by/4.0/
The gist: Recent work moves intermediate reasoning from natural-language traces into latent or cache-level representations to reduce token overhead and avoid a discrete communication bottleneck.
Terminology
Abstract
Recent work moves intermediate reasoning from natural-language traces into latent or cache-level representations to reduce token overhead and avoid a discrete communication bottleneck. However, this shift also removes a key advantage of textual reasoning: intermediate states are no longer inspectable, making it difficult to determine whether a latent state still preserves the constraints of the original query. As a result, latent reasoning typically operates in an open loop, where a latent state is produced and consumed without an input-anchored fidelity check. We propose ReLAT (Reconstruction-Guided Latent Reasoning At Test Time), a self-supervised test-time training method that closes this loop using the query itself as the reference. Our key observation is that if a latent state faithfully represents a query, the query should be recoverable from it; if the query cannot be recovered, the latent state has lost task-relevant information. ReLAT operationalizes this principle by constructing a differentiable Question -> Latent Thought -> Question cycle and optimizing query reconstruction loss through the latent thought before answer generation. This anchors opaque latent computation to the problem specification it is supposed to represent. Across mathematical reasoning, knowledge QA, and code generation benchmarks on the Qwen family, ReLAT consistently improves over single-model inference, text-based collaboration, open-loop latent collaboration, and alternative test-time training objectives. On Qwen3-8B, ReLAT raises AIME 2024 accuracy from 56.7% to 73.3%, a 16.6-point gain over the strongest open-loop latent baseline.
Sources
- Let's (not) just put things in Context: Test-Time Training for Long-Context LLMs
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- Cache-to-Cache: Direct Semantic Communication Between Large Language Models
- CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing
- Training Large Language Models to Reason in a Continuous Latent Space
- MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework
- Efficiently Learning at Test-Time: Active Fine-Tuning of LLMs
- Seek in the Dark: Reasoning via Test-Time Instance-Level Policy Gradient in Latent Space
- Language Models Are Greedy Reasoners: A Systematic Formal Analysis of Chain-of-Thought
- AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation
- SoftCoT: Soft Chain-of-Thought for Efficient Reasoning with LLMs
- You only need 4 extra tokens: Synergistic Test-time Adaptation for LLMs
- Qwen3 Technical Report
- COME: Test-time adaption by Conservatively Minimizing Entropy
- Thought Communication in Multiagent Collaboration
- Latent Collaboration in Multi-Agent Systems
Related papers
- Multi-Objective Structured Pruning of LLMs for Latency and Model Size Optimization
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Embeddings for Preferences, Not Semantics
- Can We Trust Item Response Theory for AI Evaluation?
- Diagnosing Task Insensitivity in Language Agents