acceptodds
Withdrawn from ICLR 2027

Already Solved? When to Stop and What to Return in Iterative Reasoners

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

An iterative reasoner can find a correct answer, keep computing, and ultimately return the wrong one. We study two coupled decisions: when to stop refining a state, and which generated answer to preserve. Our approach starts from a geometric observation: a hidden state can keep moving inside a region that decodes to the same answer. Margin-Motion Inclusion (MMI) compares the protection provided by the minimum logit margin with estimated remaining motion above an estimated noise floor, including a calibrated margin reserve against wrong early returns. We derive a decoding-ball guarantee without noise and conditions for recognizing the attractor's answer under persistent noise. Calibration then controls wrong early returns under exchangeability. The policy accepts and returns the first candidate that passes the test, potentially saving both iterations and restarts; otherwise, it returns the largest-margin candidate encountered. An oracle analysis separates missed solutions from delays in returning them. On recorded Sudoku and Maze trajectories from frozen Equilibrium Reasoners, fast MMI with maximum-margin fallback uses and fewer outer updates on average than terminal voting over 32 restarts of 64 updates. It also reduces errors by 20% and 75%, respectively. The gains come from recognizing and retaining answers, without changing the reasoner's learned dynamics.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
iterative reasoning, adaptive computation, latent reasoning, test-time scaling, answer selection, early-exit inference, neural dynamical systems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dRiwOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The authors explore novel halting methods for recursive reasoners. While prior work either uses a fixed number of iterations or halting based on ACT, the authors propose margin-based halting. Additionally, they propose a margin-based method of returning predictions if halting did not occur.

Critical strengths

The paper provides substantial motivational theory for their methods.

Critical weaknesses

The paper itself seems sound and provides efficiency gains but 1) the results are unconvincing, with a simple decode lock performing on-par with the proposed margin-based methods and 2) the majority of the paper is spent on motivational theory rather than experiments in actually challenging domains.

First, the paper makes several assumptions about recursive reasoners, e.g., contractive trajectories and separation assumptions. While prior theory works on similar assumptions like approaching a fixed-point, the paper here shows few convincing empirical results. Would the authors be able to provide empirical results for the separation assumption? Could we quantify the contractiveness, e.g., with an empirical contraction factor?

Additionally, I would like to see more experiments. Currently, the results might be unconvincing since the baselines in both domains are already near the results from the oracle. Seeing noticable improvements in a task with more headroom would strengthen the paper a lot. Currently, my read of the paper would suggest that the much more decode lock metric works similarly well than the proposed methods.

Questions for authors

Can we provide evidence for the assumptions? Are there results for domains where margin-based halting provides substantial improvements over decode-lock?

Reviewer_cjGzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies two inference-time decisions for iterative reasoners that expose a candidate answer after every update: when to stop refining, and which of the generated answers to return. It proposes Margin–Motion Inclusion (MMI), a stopping rule that accepts the current answer once its minimum logit margin exceeds an estimate of the remaining state motion plus a calibrated allowance, and falls back to the largest-margin candidate seen if no candidate passes. The rule is motivated by contraction-based arguments about decoding regions around attractors, and the allowance is set by split-conformal calibration. Experiments on frozen Equilibrium Reasoner (EqR) checkpoints for Sudoku-Extreme and Maze-30×30 show that MMI reaches near-oracle accuracy with far fewer outer updates than EqR's fixed-grid inference.

Critical strengths
  1. A practical and well-motivated problem. Iterative reasoners often find a correct answer before their budget ends and can later lose it. Treating "when to stop" and "what to return" as one decision is sensible, and the oracle analysis that separates missed solutions from delays in returning them (Sec. 6.3) is a useful way to show how much headroom exists.
  2. Clear empirical gains and a careful protocol. On the evaluated checkpoints, MMI matches the Sudoku oracle and comes within 0.125 points of it on Maze, while using a small fraction of EqR's 2,048-update grid. The protocol is clean: calibration and evaluation inputs are disjoint, thresholds are fixed before evaluation, results are averaged over five seeds, and simple baselines (decode locking, convergence selection, EqR's ACT head) are reported alongside.
Critical weaknesses

1. The theory adds little beyond standard results, and the rule actually used is not the one the theory covers. Lemma 1 is the standard margin–Lipschitz bound for an affine head (as the paper notes, citing Tsuzuku et al., 2018), Proposition 2 is the textbook a-posteriori error bound for a contraction mapping combined with Lemma 1, and Lemma 3 is the corresponding bound for inexact fixed-point iteration (Birken, 2015). Theorem 5 assumes that deep disagreements are rare (δsep\delta_{\text{sep}}) and concludes that accepted answers rarely disagree with the attractor, which is close to restating the assumption plus a union bound. More importantly, the practical rule in Eq. 5 departs from the theory in several ways: it uses the high-state residual ∥zt−zt−1∥\|z_t - z_{t-1}\| rather than the full-state residual the propositions require; it replaces the Lipschitz constant LL of a forward-invariant region with L^t\hat L_t, the maximum ratio of the last three step excesses, which is not an upper bound on LL; and when the previous excess is zero, any positive current excess is assigned ratio zero, i.e. treated as perfect contraction. The paper acknowledges that EqR's training "does not verify these population conditions" (App. B) and that calibration "does not verify the separation assumption of Theorem 5" (App. D), and it describes the theory as motivating the score (Sec. 3.4, App. D). The guarantee that actually holds is therefore the split-conformal one, which is standard (Lei et al., 2018; Angelopoulos & Bates, 2021). Presenting the geometric results as the basis of the method overstates their role.

2. Simple baselines already capture most of the benefit. Decode locking (stop after three identical answers) gives 98.413% at 58.27 updates on Sudoku versus 98.419% at 58.96 for MMI, and 99.35% at 41.55 updates on Maze versus 99.425% at 125.18 for MMI fast (Table 1). Calibrating the margin alone, without the motion term, gives 98.381% at 55.77 updates on Sudoku and 99.525% at 753.17 on Maze (Sec. 6.1). On Sudoku all 672 calibration scores are zero, so the allowance is zero and the fast and reserve profiles are the same rule. The headline 34.8× and 16.4× reductions are relative to EqR's full 2,048-update grid; against the convergence selector on Sudoku, the error reduction is 0.8% (Sec. 1). Returning the largest-margin intermediate candidate is also close to the maximum-confidence selection of Bansal et al. (2022). Taken together, it is unclear what MMI adds over confidence- or patience-based early exit.

3. Only one model is evaluated, while the claims are about iterative reasoners in general. All experiments use EqR checkpoints on two puzzle tasks. Sec. 2 introduces HRM and TRM as instantiations of the same framework, but App. F states that "our experiments concern EqR, with HRM and TRM providing architectural context", and Sec. 7 leaves "transfer beyond these checkpoints" open. Nothing in MMI is specific to EqR: margins, residuals and decode locking are available for HRM and TRM, whose deterministic updates would also test the noise-free case of the theory directly. Since evaluation is offline replay of recorded trajectories and needs no training, adding the public HRM and TRM checkpoints would be inexpensive. Without them, the title's claim about "iterative reasoners" is not supported. The paper also reports update counts rather than wall-clock time, although saving computation is its main motivation.

Questions for authors
  1. How does MMI compare with decode locking and margin-only calibration on HRM and TRM, using their released checkpoints? Does the noise-free guarantee (Proposition 2) hold empirically there?
  2. How often does L^t\hat L_t underestimate the actual contraction behaviour of EqR, for example compared with a Jacobian-based or longer-window estimate? Why is a positive excess after a zero excess assigned ratio zero?
  3. What does MMI provide over decode locking that would justify its additional complexity, given the results in Table 1?
  4. What are the wall-clock savings of MMI over the grid and over decode locking, including the cost of computing margins and residuals at every update?
Reviewer_p2dAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Margin-Motion Inclusion (MMI), a method that jointly addresses early stopping and answer selection in iterative reasoning models. By combining logit margins, estimated remaining state motion, and statistical calibration, MMI aims to reduce unnecessary computation while preserving correct intermediate answers. Experiments on Sudoku and Maze show promising improvements in accuracy and computational efficiency, although the method’s generalizability beyond EqR and its practical inference speedup remain unclear.

Critical strengths
  • The proposed method has theoretical motivation and shows promising results on Sudoku and Maze tasks, reducing computational updates while maintaining or improving accuracy.
Critical weaknesses
  • Experiments are restricted to a single model architecture (EqR) and two structured tasks. It remains unclear whether MMI generalizes to other reasoning models or more complex tasks.

  • MMI provides only marginal accuracy improvements over Decode Lock, with gains of 0.006 percentage points on Sudoku and 0.075 percentage points on Maze for MMI Fast, while requiring slightly more and substantially more updates, respectively.

  • Although the paper reports results across five random seeds, it does not provide standard deviations, confidence intervals, or statistical significance tests. It is therefore unclear whether the reported improvements are robust or statistically meaningful.

Questions for authors
  • MMI Fast achieves only marginal accuracy improvements over Decode Lock (0.006 percentage points on Sudoku and 0.075 percentage points on Maze), while requiring more updates. Although experiments use five random seeds, no variance estimates or confidence intervals are reported. Are these improvements statistically significant and consistent across seeds?

  • How much of MMI’s performance comes from its stopping criterion versus its maximum-margin answer selection strategy? Could the authors provide ablations and comparisons against maximum-margin selection alone, as well as confidence-based and stability-based stopping baselines, under matched computational budgets?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.