- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes Energy-guided Recursive Model (ERM), an inference framework for fixed iterative models. On puzzles, it scores candidate outputs with soft retrieval over valid local patterns and explicit global constraints. It uses these scores for selection, parallel-tempering exchanges, and a batch-level stopping rule. On language modeling, it combines candidate probabilities with retrieved continuation statistics through a normalized geometric mean. ERM matches the candidate-pool oracle on the three main puzzle experiments. Parallel tempering adds three Sudoku solves. On a RedPajama-V2 proxy, the language correction reduces perplexity from 44.189 to 43.419. The paper also reports rule-checker controls, memory ablations, and reduced-depth Maze inference.
Critical strengths
The shared-pool evaluation separates candidate generation from candidate selection. Table 1 measures how much each selector recovers from the same available answers. ERM improves PPBench from 251 to 265 solves relative to q-head selection. It also recovers every available solution in the main Sudoku and Maze pools. Reporting the oracle prevents these selection gains from being confused with generating new solutions. Tables 9 and 10 provide valuable additional controls using simpler rule scores.
Critical weaknesses
The central reasoning results do not establish a benefit specific to Hopfield retrieval. In Table 9, global-only scoring, unweighted violation count, first-valid selection, and full ERM all solve exactly 993/1000 Maze instances. In Table 10, a discrete score using the complete local and global rules solves the same 265/301 PPBench instances as full ERM. Thus, these controls already explain the headline selection benefit through ordinary task verification. The local-versus-global ablations show that all relevant rules matter. They do not demonstrate that the soft memory formulation adds value over evaluating those rules directly. The claimed broader role in sampling and stopping could distinguish the method, but this is not established against checker-based PT or stopping. PT adds only three Sudoku solves, without repeated seeds. Table 4 compares different candidate pools, and stable batch energy is not a correctness certificate. The evidence is useful, but it supports a narrower contribution than the main Hopfield-guidance claim.
The language result and the claim of marginal overhead need a cleaner evaluation boundary. Equations 9-10 reduce exactly to p_ERM(v) proportional to sqrt(p0(v) * ((1-g_t)p0(v) + g_t q_t(v))). This is a memory-based probability correction. The supplied mixture and memory-form ablations are useful, but they do not establish a special connection to recursive reasoning. Training-free memory augmentation for language prediction already has a direct precedent in Grave et al., "Improving Neural Language Models with a Continuous Cache," ICLR 2017: https://openreview.net/pdf?id=B184E5qee . More importantly, Appendix B.2 states that the same 33k-document proxy supports development and evaluation. The reported bootstrap interval measures document-sampling uncertainty within that proxy. It does not account for model or memory choices made using it. Finally, the 2.83-hour language runtime covers the full comparison, and the development timing covers all branches together. Table 13 excludes retrieval and constraint scoring. These records do not isolate the added cost required to substantiate "marginal inference overhead."
Questions for authors
-
On matched candidate pools and repeated seeds, where does the soft Hopfield score improve over the complete discrete verifier? Please include checker-based PT or stopping if those are the intended advantages beyond final selection.
-
Which documents and results informed memory, gate, checkpoint, and variant selection? Please report the frozen final configuration on a separate evaluation split that was not used for development. Does the gain persist with one candidate, or is there evidence that recurrence or multiple trajectories are important to the correction?
-
Please separate candidate-generation time, memory construction, lookup, constraint scoring, normalization, and peak memory. What is the measured incremental cost over candidate averaging and over a single candidate? For Maze stopping, what is the wall-clock comparison with matched PT fixed-depth and first-valid stopping?
-
What exactly are the PTRM and DeepSeek-V4 bars in Figure 1? Please provide their generator identities, prompts, candidate budgets, and evaluation protocol, and explain how these bars relate to the shared EqR selectors in Table 1.