Uncovering Latent Reasoning Strategies in Language Models

Hugging Face Daily Papers Papers
language-models reasoning latent-variable variational-inference posterior-collapse representation-learning

Summary

This paper proposes a method to decompose the response distribution of language models into structured, strategy-conditioned representations using a latent variable, addressing posterior collapse with a model-directed reconstruction objective.

A language model p_θ(y mid x) trained on reasoning tasks learns to solve problems via multiple distinct strategies, yet these strategies are implicit and entangled within the model's response distribution. We study the problem of decomposing the response distribution of a given pretrained language model into a structured, strategy-conditioned representation. Specifically, we learn a latent-variable factorization p_θ(y mid x) leadsto (r_ϕ(z mid x), g_ϕ(y mid x,z)), where a router r maps each input to a distribution over latent strategies z and a generator g produces the response conditioned on that strategy. A key challenge is that the generator, initialized from the base model, already represents p_θ(y mid x) without using z. Standard variational inference therefore gives the model no incentive to route information through z and can yield a severe form of posterior collapse. To address this, we propose a variational objective that measures fractional information gain relative to the base model's response loss and concentrates reconstruction pressure on tokens with high base model surprisal, encouraging z to encode strategy-relevant response variation. We introduce a benchmark of multi-strategy algorithmic tasks and show that this objective recovers latent codes aligned with distinct reference strategies while preserving the base model's response distribution.
Original Article
View Cached Full Text

Cached at: 07/29/26, 07:54 PM

Paper page - Uncovering Latent Reasoning Strategies in Language Models

Source: https://huggingface.co/papers/2607.17674 A language model trained on reasoning tasks learns a response distribution that mixes several distinct solution strategies, but these strategies remain implicit and entangled within the model’s response distribution, with no explicit variable for identifying or controlling the strategy used during generation.

We study how to decompose this distribution into a structured, strategy-conditioned representation in which a latent variable controls the generation strategy. The key challenge is that the generator already represents the target response distribution at initialization without using the latent, leading to a severe form of posterior collapse in standard variational inference.

We address this with a model-directed reconstruction objective that focuses learning on strategy-relevant response variation. On a controlled benchmark, our method recovers latents aligned with reference strategies and consistent across inputs while preserving the base model’s response distribution.

Similar Articles

Large Reasoning Models Are (Not Yet) Multilingual Latent Reasoners

arXiv cs.CL

This paper investigates multilingual latent reasoning in large reasoning models across 11 languages, revealing that while latent reasoning capabilities exist, they are unevenly distributed—stronger in resource-rich languages and weaker in low-resource ones. The study finds that despite surface-level differences, the internal reasoning mechanisms are largely aligned with an English-centered pathway.

Decomposing and Steering Functional Metacognition in Large Language Models

arXiv cs.CL

This research paper investigates functional metacognition in Large Language Models, demonstrating that internal states like evaluation awareness and self-assessed capability are linearly decodable from residual stream activations. The authors propose a mechanistic framework to steer these states, showing causal control over reasoning behaviors, verbosity, and safety responses.