Invariant-Measure Reasoners: Stable Representations for Latent Reasoning

arXiv:2610.10996 · cs.LG · Submitted 2026-10-07 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Invariant-Measure Reasoners: Stable Representations for Latent Reasoning".

Jane: The gist:

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So, let's talk about the title and who wrote this stuff. We have "Invariant-Measure Reasoners: Stable Representations for Latent Reasoning" by Yuto Inui from The University of Osaka, Takuya Konishi also from there, and Yoshinobu Kawahara from The University of Osaka and RIKEN.

Jane: It sounds a bit technical, but the core idea is about using an invariant measure as a stable representation for these latent states. Think of it as finding the long-term probability distribution of where those states end up after many updates.

Lu: They are introducing this invariant measure because it’s preserved by the recurrent updates, which is a key mathematical property from the work of Eckmann and Ruelle in one thousand nine hundred eighty-five <ref:2610.10996#pg1>.

Meng: So, instead of looking at every single state in the sequence, they want to summarize that long-term behavior into one stable distribution. That sounds like a way to get ahead of the instability issues we see with recurrent depth increasing.

Lalam: It’s about replacing the unpredictable path of individual states with a predictable distribution that stays the same regardless of how deep you go.

The paper's summary: Tom: The main takeaway from this paper, "Invariant-Measure Reasoners: Stable Representations for Latent Reasoning," is that we can predict from the expectation of the prediction head’s output under this invariant measure.

Jane: Instead of just using the current latent state, you take the average over all possible states according to that invariant measure. This averaging process, which they call z-star, makes the resulting prediction stable across different recurrent depths.

Lu: The paper explains that if a trajectory converges to some compact set—say a bounded region—then this invariant measure is supported on that set of states.

Meng: They use an empirical measure, called mu-hat T, at a finite stopping depth T as an approximation for the true invariant measure. So they calculate z-hat T as the expectation over that approximation.

Lalam: This averaging helps reduce sensitivity to how deep you go in your recurrent updates and improves robustness in terms of prediction accuracy across different depths.

The paper's improvements: Tom: The paper suggests a few ways to actually use this framework. First, they look at two training approaches: fine-tuning only the prediction head, or training the whole model from scratch using a deep supervision procedure.

Jane: For fine-tuning, they treat the latent states as constants when calculating the loss for just the prediction head's gradients, which means you don't have to backpropagate through all those recurrent updates.

Lu: For training from scratch, they combine ImR with a deep supervision setup where an inner loop does L recurrent updates and an outer loop computes the loss from that averaged logit.

Meng: That deep supervision approach is interesting because it lets the model learn latent dynamics that are specifically suited for prediction using this invariant measure representation.

Lalam: It’s about training a model that learns to be stable from the start, rather than just hoping it settles into a fixed point eventually.

Conclusion: Tom: So, to wrap up on "Invariant-Measure Reasoners: Stable Representations for Latent Reasoning," the main point is that using an invariant measure provides a stable representation for latent dynamics and predicting from its expectation stabilizes predictions across recurrent depths.

Jane: This means we don't necessarily need every single latent state to settle down into one specific fixed point to get reliable answers, which is a big conceptual shift for how we think about these models.

Lu: The framework suggests that we can leverage dynamics that might seem unstable in isolation but are stable when viewed through the lens of this invariant measure.

Meng: From an engineering standpoint, the averaging provides a practical way to get robust outputs without needing extremely deep, potentially infinite, recurrent runs just to be sure.

Lalam: It’s about creating a more reliable way for latent reasoning models to function in real-world tasks like solving mazes or Sudoku by focusing on the overall distribution rather than single state trajectories.

Yuto Inui, Takuya Konishi, Yoshinobu Kawahara

The University of Osaka · RIKEN

cs.LG

Submitted: 2026-10-07

Updated: 2026-10-07

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 92/100

The gist: The gist: Invariant-Measure Reasoners (ImR) introduce a framework that uses an invariant measure as a stable representation of latent dynamics to predict from the expectation of the prediction

Key concepts

Invariant Measure: An invariant measure is a probability distribution over latent states that remains unchanged when subjected to recurrent updates by the model. It captures the long-term, stable behavior of the latent trajectory, even if individual states do not settle on a single fixed point.
Prediction from Invariant Measure: Instead of predicting from one specific state, ImR predicts using the expectation (average) of the prediction head's output calculated over all states weighted by the invariant measure. This averaging process makes the prediction robust against changes in recurrent depth or trajectory convergence.
Prediction Instability: This occurs when latent states repeatedly change even after many recurrent updates, leading to inconsistent predictions across different depths of the model. ImR mitigates this instability by using a stable, measure-based representation instead of relying on an unstable single latent state.

Terminology

Latent Reasoning — 1× in this paper · explained in 3 episodes

Summary

The gist: Invariant-Measure Reasoners (ImR) introduce a framework that uses an invariant measure as a stable representation of latent dynamics to predict from the expectation of the prediction head’s output under this measure, which reduces prediction instability and improves accuracy in many settings on maze and Sudoku tasks.

Introduction and Motivation

Latent reasoning models repeatedly update a latent state using the same recurrent block, but as recurrent depth increases, the sequence of latent states may converge to a compact subset of the state space without necessarily converging to a fixed point <ref:2610.10996#pg6>. Existing models typically predict by applying a prediction head to a single latent state, but this can lead to prediction instability across recurrent depths <ref:2610.10996#pg5>. This instability is illustrated by trajectories that approach a compact subset of the state space without converging to a single fixed point, which can cause predictions to alternate between correct and incorrect solutions <ref:2610.10996#pg4>.

Invariant Measures of Latent Dynamics

The paper defines an invariant measure as a probability distribution over latent states that is preserved by recurrent updates <ref:2610.10996#pg5>. This property is expressed using the pushforward operator, where for a probability measure µ, f µ is the distribution obtained by applying f to states distributed according to µ <ref:2610.10996#pg4>. An invariant measure satisfies the condition f µ = µ <ref:2610.10996#pg5>. The support of this measure reflects the long-term behavior of the latent trajectory, and if the trajectory converges to a compact set B, then µ⋆ is supported on B <ref:2610.10996#pg4>.

Prediction from an Invariant Measure

ImR predicts from the expectation of the logits under this invariant measure, defined as z⋆ = Es∼µ⋆ [g(s)] = Z S g(s) dµ⋆ (s) <ref:2610.10996#pg5>. This expected output is invariant under recurrent updates because Es∼µ⋆ [g(f(s))] = Es∼f µ⋆ [g(s)] = z⋆ <ref:2610.10996#pg5>. The invariant measure µ⋆ can be approximated using the empirical measure µˆT at a finite stopping depth T, resulting in the prediction zˆT = Es∼µˆT [g(s)] <ref:2610.10996#pg5>. This averaging reduces sensitivity to the stopping depth T and improves prediction robustness across recurrent depths <ref:2610.10996#pg5>.

Training with Prediction from an Invariant Measure

The paper proposes two training approaches: fine-tuning only the prediction head or training models from scratch <ref:2610.10996#pg5>. For fine-tuning, the loss is computed from the averaged logits in equation 6, treating latent states as constants so that gradients pass only through g <ref:2610.10996#pg5>. For training from scratch, ImR is combined with a deep supervision procedure where an inner loop performs L recurrent updates and an outer loop computes the loss from the averaged logits <ref:2610.10996#pg5>. Truncated backpropagation through time (TBPTT) is used to limit gradient computation cost <ref:2610.10996#pg5>.

Experimental Results

Experiments on maze and Sudoku tasks showed reduced prediction instability and improved accuracy in many settings <ref:2610.10996#pg4>. For EqR trained from scratch, ImR achieves higher accuracy than the released EqR model on both datasets at large recurrent depths <ref:2610.10996#pg5>. Furthermore, ImR (FT) reduces prediction instability in the non-fixed group and maintains high accuracy even for trajectories classified as non-fixed <ref:2610.10996#pg5>. The results suggest that ImR can leverage otherwise destabilizing dynamics for latent reasoning <ref:2610.10996#pg4>.

Conclusion

ImR successfully addresses prediction instability at large recurrent depths in latent reasoning models by using an invariant measure as a stable representation and predicting from the expectation of the prediction head’s output under this measure <ref:2610.10996#pg5>. These findings support a broader formulation of latent reasoning based on convergence of empirical measures, allowing stable prediction without requiring individual states to converge to a fixed point <ref:2610.10996#pg4>.

AI Use Statement

We used generative AI tools to assist with drafting and revising the manuscript, searching for relevant literature, and discussing mathematical formulations and their assumptions <ref:2610.10996#pg5>. We also used these tools to write code for the numerical experiments and to discuss the interpretation and presentation of experimental results <ref:2610.10996#pg5>. The authors reviewed and revised the AI-assisted material, including the generated code, and take full responsibility for the final content of this work <ref:2610.10996#pg5>.

Reproducibility Statement

Section 3 describes the prediction rule and training procedures for ImR, with an overview of training from scratch in Algorithm 1 <ref:2610.10996#pg5>. Appendix B.1 describes the datasets, their preparation, and the training, validation, and test splits <ref:2610.10996#pg5>. Appendix B.2 specifies the model architectures, while Appendix B.3 details the training and evaluation settings and the procedures for selecting hyperparameters and checkpoints <ref:2610.10996#pg5>. Appendix B.4 provides the criteria for classifying latent trajectories and the procedures for calculating prediction stability proportions within each group <ref:2610.10996#pg5>. Appendix B.5 describes how the PCA projection and prediction regions in Figure 1(a) were constructed and reports the corresponding predictions from the actual latent states <ref:2610.10996#pg5>.

Acknowledgments

This work was supported by JST BOOST Grant Number JPMJBS2402, JSPS KAKENHI Grant Numbers JP25K1523 and JP26H02498, and JST Moonshot R&D Program Grant Number JPMJMS263K <ref:2610.10996#pg5>.

References

Cem Anil, Ashwini Pokle, Kaiqu Liang, Johannes Treutlein, Yuhuai Wu, Shaojie Bai, J. Zico Kolter, and Roger B. Grosse. Path independent equilibrium models can better exploit test-time computation <ref:2610.10996#pg5>. Junyeob Baek, Mingyu Jo, Minsu Kim, Mengye Ren, Yoshua Bengio, and Sungjin Ahn. Generative recursive reasoning <ref:2610.10996#pg5>. Shaojie Bai, J. Zico Kolter, and Vladlen Koltun. Deep equilibrium models <ref:2610.10996#pg4>. Shaojie Bai, Vladlen Koltun, and J. Zico Kolter. Stabilizing equilibrium models by Jacobian regularization <ref:2610.10996#pg4>. Andrea Banino, Jan Balaguer, and Charles Blundell. PonderNet: Learning to ponder <ref:2610.10996#pg4>. Arpit Bansal, Avi Schwarzschild, Eitan Borgnia, Zeyad Emam, Furong Huang, Micah Goldblum, and Tom Goldstein. End-to-end algorithm synthesis with recurrent networks: Extrapolation without overthinking <ref:2610.10996#pg5>. Jeremy Bernstein, Yu-Xiang Wang, Kamyar Azizzadenesheli, and Animashree Anandkumar. signSGD: Compressed optimisation for non-convex problems <ref:2610.10996#pg5>.

References (Continued)

J.-P. Eckmann and D. Ruelle. Ergodic theory of chaos and strange attractors <ref:2610.10996#pg5>. Katie E. Everett, Lechao Xiao, Mitchell Wortsman, Alexander A. Alemi, Roman Novak, Peter J. Liu, Izzeddin Gur, Jascha Sohl-Dickstein, Leslie Pack Kaelbling, Jaehoon Lee, and Jeffrey Pennington. Scaling exponents across parameterizations and optimizers <ref:2610.10996#pg5>. Jonas Geiping, Sean Michael McLeish, Neel Jain, John Kirchenbauer, Siddharth Singh, Brian R. Bartoldson, Bhavya Kailkhura, Abhinav Bhatele, and Tom Goldstein. Scaling up test-time compute with latent reasoning: A recurrent depth approach <ref:2610.10996#pg5>. Alex Graves. Adaptive computation time for recurrent neural networks <ref:2610.10996#pg5>. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F.

Improvements for AI systems

Bold Header: Invariant Measure Prediction

This system can predict from the expectation of the prediction head’s output under this measure (Equation 5), which is approximated by the averaged logits zˆT (Equation 6). This stabilizes predictions even when individual latent states do not converge to a fixed point, addressing the instability seen in existing models where trajectories repeatedly trace a loop reminiscent of a limit cycle.

Bold Header: Fine-Tuning for Stability

The system can be fine-tuned by fine-tuning only the prediction head while treating latent states as constants, aiming to reduce prediction instability in an existing model without backpropagating through recurrent updates. This allows existing models to achieve stable predictions and improve accuracy even with such behavior, as shown in Table 1.

Bold Header: Training from Scratch with Deep Supervision

The system can be trained jointly by combining ImR with deep supervision, where the loss is computed from the averaged logits zˆL (Algorithm 1). This allows the model to learn latent dynamics suited to prediction with ImR, leading to higher accuracy than models trained from scratch using only the final state prediction.

Bold Header: Generalization Across Latent Space

The framework enables better generalization by using an invariant measure as a stable representation of latent dynamics (Section 3), which is preserved under recurrent updates, suggesting that ImR can leverage otherwise destabilizing dynamics for latent reasoning.

Bold Header: Handling Non-Fixed Point Behavior

The system can achieve high accuracy even when exhibiting non-fixed-point behavior, as the paper shows models trained with ImR exhibit non-fixed-point behavior more frequently than existing models yet achieve high accuracy even with such behavior.