- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes REST (Representation-Supervised Thoughts), a training framework for latent recursive single- and multi-agent LLM systems. With the base models and inner links frozen, REST trains inter-agent or inter-round connectors using auxiliary objectives motivated by four properties: causality, minimality, separability, and stability. The authors argue that answer-level cross-entropy alone insufficiently constrains latent representations and provide theoretical analyses connecting these properties to prediction quality. Experiments across mathematics, science, medicine, and code generation report improvements over CE-only training and adapted representation-supervision baselines, accompanied by analyses of latent representations and generation behavior.
Critical strengths
S1. The figures are clear and well designed, making the system architecture and the proposed auxiliary objectives easy to understand.
S2. The method requires no architectural changes or additional parameters at inference, making it straightforward to integrate into the latent recursive systems studied.
S3. The evaluation covers multiple task domains, model scales, and both single- and multi-agent settings. The appendix also reports extensive hyperparameter sweeps and negative results, providing a useful view of the method’s limitations.
Critical weaknesses
W1. The applicability of the Thought Collision analysis to cross-question comparisons is unclear. L.167–171 claim that colliding thoughts force the consumer to produce nearly identical answer distributions, with complete collapse leading to equal probability for the two targets. However, Definition 2.1 (L.124–128) and L.142–144 explicitly state that the consumer conditions on both its own context and the transferred thought. Thus, even with fixed consumer parameters, , because the question contexts may differ. The derivation relies on Assumption A.8 (L.1631–1639, Eq. 56), but the paper does not establish why output-distribution differences across distinct contexts can be bounded solely by thought distance. The authors should clarify the scope of this assumption and use interventions that hold the context fixed while swapping or replacing thoughts to test whether the clustering observed in Figure 6 actually affects answer prediction.
W2. The Stability analysis is not directly connected to the implemented training objective. Equation (6) only requires predicting the producer’s output entropy from the thought. Entropy is a scalar summary, and different distributions can have identical entropy. Propositions A.18–A.19 characterize the expected log-loss advantage of mixture prediction over a single sampled transfer; Assumption A.9 and Corollary A.9 establish that a latent transfer attains this advantage when it realizes the mixture distribution. However, the paper does not establish why optimizing entropy regression would move the consumer toward that mixture. The analysis therefore motivates the potential value of preserving distributional information, but provides only indirect support for the implemented surrogate. The authors should distinguish the conditional theoretical result from the heuristic motivation and empirically test whether entropy supervision improves the consumer’s use of producer uncertainty.
W3. Computational-budget matching is insufficiently documented. The paper emphasizes matched compute, but Appendix B.2–B.3 primarily specifies shared training steps and batch size, without reporting actual training FLOPs or runtime for REST and auxiliary-supervision baselines such as SIM-CoT. Their additional computation differs: REST’s causality and minimality objectives involve additional prediction or reconstruction computations, whereas the adapted SIM-CoT baseline trains auxiliary decoders at each stage. Equal training steps therefore do not establish equal computational cost. At inference, the decoded-token counts in Table 4 also do not capture the full cost of latent computation. Reporting training FLOPs, runtime, and end-to-end inference costs would make the accuracy–cost trade-offs and comparison fairness easier to assess.
Questions for authors
Q1. How are the auxiliary-loss weights and recursion rounds selected? L.295–297 state that each system is reported at its best-performing recursion round, while the Table 2 caption (L.302) reports each row at its best ; the search grids appear in L.1810–1814. Are these choices made using a held-out validation set independent of the reported benchmarks? Please clarify the selection criterion and whether all baselines follow the same protocol.
Q2. How sensitive are the gains to the latent thought budget and the training–inference length mismatch? L.1815–1816 specify thoughts of up to 80 positions during training and a fixed 32 latent steps at inference. Figure 12 provides a budget sweep, but does it vary only inference length or retrain models for each budget? Do the gains persist when training and inference lengths are matched?
Q3. Why is the input-suppression term disabled when combining properties? L.1811–1812 specify for standalone minimality but for composition. Consequently, the minimality component in “Best pair” and “All properties” reduces to output reconstruction. What motivates this change, and how do the combined objectives perform when is retained?
Other questions see weakness.