acceptodds
Under review as a conference paper at ICLR 2027

Same Readout, Different Rollout: When to Let Go

Abstract

Two histories can reach the same complete state of a modeled system and yield identical, correct model readouts, yet the model predicts different futures under the same actions. We study such pairs after a memory value is deleted or an object returns by a different path. In the tested world-model conditions, every initially correct pair agrees after one action, yet 8.2–16.8% diverge on at least one of three fixed 16-action continuations; at the first split, one branch is usually correct. We test the classical readout-closure requirement on these pairs: an update must preserve the equivalence defined by current answers. Reconstructing a state from those answers restores pair agreement and increases the fraction of world-model pairs whose two branches predict every tested 16-step endpoint correctly. Memory results show why agreement and accuracy remain separate. A structured updater illustrates how identifier-renaming and storage-reordering symmetries can be preserved through future updates. Correct current answers therefore need not make learned states interchangeable for continued computation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation learning, state abstraction, behavioral equivalence, world models, neural memory, long-horizon prediction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_PH9iOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper asks whether learned models that give identical, correct answers about a system's current state stay interchangeable after further updates. The authors build pairs of histories that reach the same true state and apply the same actions to both. Predictions often diverge later. Rebuilding the state from its current answers, a classical requirement called readout closure, restores agreement and makes the world model nearly perfectly accurate, though in memory agreement doesn't guarantee accuracy.

Critical strengths
  • The phenomenon is concrete and clearly demonstrated. The paired design holds the checkpoint, true state and continuation fixed, which isolates the effect of history.
  • The reporting is transparent. It includes mixed and negative results, prespecified success criteria, and a failed earlier protocol.
  • Reconstruction gives large and consistent accuracy gains in C-SWM.
  • NRT comes with a formal equivariance guarantee and proof.
Critical weaknesses
  • Lack of clarity and readability. The abstract and introduction uses terms like “readout” and “retained state” without explaining what they mean. The NRT (Section 4 onwards) only loosely ties to the rest of the paper so it reads like two papers. There is not enough coherence to make this paper publishable.
  • Limited experimental results. Small number of samples are used with only 30-53 pairs for memory experiments. Two memory checkpoints are not independent models. Arbitrary toy tasks only with a 64-slot table and 5x5 grid with five objects.
  • Lack of generalizability. It is unclear how it would generalize to real world models where the full state cannot be fully read out.
  • NRT is not better than Structured SIT. A simpler Structured SIT beats NRT on accuracy on all the conditions, which makes it unclear how NRT is a contribution.
  • Lack of novelty and incremental contribution. The paper acknowledges that its closure criterion is the classical congruence/quotient condition from universal algebra. The contribution therefore rests mainly on the experiments, whose scope is limited.
Questions for authors
  • What is the unifying claim connecting Sections 2–3 with Section 4?
  • Do the results hold with independently trained memory models and more C-SWM seeds?
  • How would reconstruction work when the readout is partial or the state cannot be fully decoded?
  • Given that Structured SIT matches NRT’s guarantee with higher accuracy, what is the case for NRT?
1 comment

Thank you for asking how the two parts connect. The revision clarifies the relation that should survive updates in each setting and adds tests across model initializations and partial readouts.

From observed failure to a design specification. In the original §§2–3, two histories can reach the same sufficient source state and give the same correct current answers, yet yield different predictions under identical actions. One C-SWM reconstruction from each state's own answers preserves those answers and improves future source correctness with the same updater (original Tables 3–4, revised Figure 3). The revised Introduction calls this an interchangeability failure. Section 4 treats the decoded-store setting, where we can specify exactly how a change of state encoding and its corresponding actions should transport the predictions.

Independently trained models. We added three C-SWM initializations on the original training data and one independent memory-model initialization. Two grid initializations diverged and benefited from reconstruction in both history families. The third was already correct at all tested endpoints. The independently initialized memory model diverged after both deletes and overwrites (§3, revised Tables 4–5, Appendices B.2 and B.5).

Partial readouts. We added a check with dialogue trackers factorized by restaurant attributes (WOZ 2.0). With fixed weights and deterministic updates, replaying the observed prefix from the fixed initial state recovers the 96 of 840 recurrent coordinates that preserve the tracker's own area and price predictions under any common continuation. This identifies a compact history-dependent state for those two predictions (revised Appendix B.9).

Purpose of NRT and Structured SIT. NRT realizes the specified correspondence by design. Transporting the decoded store and its action sequence together under identifier and slot permutations preserves corresponding predictions for finite source-legal rollouts in deterministic, exact arithmetic (Theorem 1). We revised §4.2 to separate the original slot-augmented SIT control from the direct comparison with two alternative architectures on the same interface. Structured SIT checks the same correspondence and accuracy criteria in another architecture. It leads on both accuracy measures in all four conditions. NRT comes within 1.2 percentage points of it in mean query accuracy in each condition and within 2.1 points in full-state accuracy for tiered memory with 1,024 training transitions. The full-state gap is larger for Records / 1,024 (original Table 7, revised Appendix E.5, Table 10).

Reviewer_s1nXOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

Authors argue that world models are, in some cases, path dependent: given two sequences {s1, ...} and {t1, ...} which result in a shared end state s_n = t_m (the "readout"), learned world models should treat the generating sequences as interchangable if they depend only on the current state; authors provide a model to account for when this may or may not happen, and provide empirical evidence that in some cases, agreement of readouts is insufficient guarantee that learned models will not diverge in their representations at future timesteps.

Critical strengths

Authors present a thorough formalization of properties of memorylessness which are relevant to describing some world models.

Critical weaknesses

I find this paper very hard to understand, for two reasons. First, the authors do not make a strong attempt to present their findings in a way which is amenable to contextualization and evaluation: there is little attempt to connect the main ideas of the first two sections to prior literature, nor is the motivation for the class of problems studied clearly laid out. The authors do not specifically state a hypothesis for how learned world models ought to behave, nor is there much if any discussion of the impacts of the work On the whole, the paper is written in a way which impedes quick understanding of the ideas presented. Second, it's not obvious to me why the main property discussed (that learned world models diverge given different inputs with identical readouts) is important, relevant, or a right/wrong model for real-world systems. The main claim seems to be that learned world models tend to not be memoryless, and rather depend on prior states to inform future computation. There are some cases when that is bad (eg, if the underlying process is known to be Markov) and others where it is good (eg, many real-world timeseries problems explicitly require memory). It's very unclear to me what is at stake in noting that trained world models pattern like one of these or the other. Based on this, it's difficult to evaluate the impact of the paper.

Questions for authors

Could the authors please address the following points:

  • what are the central uses for the various world models discussed here, and how do they relate to the learned neural model used in the experiments?
  • when should we expect systems to be "memoryless" or otherwise?
  • what is at stake if the learned neural model exhibits preferences for memory behaviors?
  • what are the key implications of the results of this work?
3 comments

Thank you for asking when history is useful. We revised the motivation and Discussion to show when histories with the same current answers should give the same future predictions and how to check that in a learned model.

When history should matter. The revised Introduction distinguishes the sufficient source state from a model's retained state and the answers it reads out. Earlier positions can help when velocity is unobserved. In our grid the full positions determine what a given move does, so two correct readouts of the same positions should support the same prediction under that move (§1).

Prediction consequences. In the original object-based grid model (C-SWM), 43/256 pairs in one condition had identical, correct positions initially and separated within 16 shared actions. At the selected first split, exactly one branch was source-correct in 40/43 cases (original Table 5, revised Appendix B.7, Table 8). Rebuilding each state once from its own positions, with the same updater and no new observation, raised strict three-endpoint correctness from 197/256 to 254/256. In another condition, 36 of the 52 pairs improved by reconstruction had agreed throughout (original Tables 3–4, revised Figure 3 and Tables 6–7). This is why the revised Discussion recommends checking current-answer preservation and scoring future predictions against the source under the same updater and actions (§6).

How the models are used. C-SWM predicts positions for candidate-action rollouts. The revised Introduction explains why keeping those predictions equal gives the same candidate score under a shared goal and fixed scoring rule. The memory predicts key values after writes, copies and deletions. NRT addresses a decoded-store interface, where transforming an encoding together with its actions preserves corresponding predictions under the stated guarantee (§4, Theorem 1). We added an offline example in which replaying the same later utterances from the earlier rather than the later of two turns with the same correct food, area, and price answers can change final restaurant suitability (§3, Appendix B.9).

Relation to prior work. Revised §5 connects classical congruence to state abstraction and predictive representations, including Givan et al. (2003) and Castro et al. (2009). It also discusses Vafa et al.'s tests of histories reaching the same automaton state and Fathi et al.'s use of re-encoding. Our diagnostic holds the sufficient source state, current correctness, learned updater, and continuation fixed.

I don’t think that these posts adequately address my main concerns. It is still unclear, for instance, why the we should want models to be memoryless, or what the main implications of the findings here are.

I maintain my score in light of this.

Thank you for sharing your confusion. In short we study the phenomenon that models holding memory increases instability and encourages drift of the models' output, which leads to errors.

We hope this clarification helps.

Reviewer_KJcMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies whether two learned states with identical and correct current readouts remain equivalent under future updates. It shows that different histories can produce the same source state and current answers, yet lead to divergent future predictions due to differences retained in the internal representations. The authors formalize this as a failure of readout closure using the classical notion of congruence, and demonstrate it in neural memory models and contrastive structured world models. They find substantial delayed divergence. They then propose state reconstruction to restore agreement and introduce the Nominal Replacement Transducer (NRT) to preserve identifier and storage-slot covariance.

Critical strengths

The main observation is interesting and relevant. Two learned states can produce identical, fully correct readouts while retaining different information about their histories, which can influence subsequent predictions. This is an important distinction, particularly for recurrent memory systems and world models that repeatedly update their internal representations.

The paper provides a reasonably clean mathematical characterization of when two states with identical readouts remain equivalent under arbitrary finite action sequences. The linear specialization also helps illustrate how hidden differences can become visible through subsequent updates.

The distinction between agreement and source correctness is important. Reconstructing a common state guarantees agreement between initially equivalent states, but does not automatically guarantee that the resulting predictions are correct. The paper makes this distinction explicit both theoretically and experimentally.

Critical weaknesses

The paper is overall very badly written with substantial AI use, as evident from the sloppy writing in almost all places. The overall novelty of the paper is somewhat limited by the fact that its main theoretical contribution is largely a restatement and specialization of classical congruence and quotient constructions. The equivalence between closure under updates and the existence of a well-defined update on the quotient space is mathematically natural. While applying this perspective to learned representations is interesting, I am not fully convinced that the theoretical results themselves constitute a sufficiently substantial contribution for ICLR.

The empirical evaluation is relatively narrow. The history-pair experiments use a small number of model checkpoints and highly controlled environments. In particular, the C-SWM experiments involve only five objects on a 5 × 5 grid without inertia, while the neural memory experiments use a relatively small number of initially correct pairs (30–53 per condition). These settings are useful for isolating the phenomenon, but it is unclear how frequently the same problem occurs in more realistic world models, larger memory systems, or modern agent architectures. The practical significance of experimental evidence for delayed separation is not fully established. The paper constructs pairs of histories that lead to identical source states and then examines their future behavior. However, it remains unclear how often such pairs arise naturally during ordinary model operation, or whether the observed divergence significantly affects downstream tasks beyond the controlled environments considered here.

Questions for authors

How frequently do these history-dependent divergences occur in more realistic settings? Can you provide evidence from larger world models, recurrent language models, or practical agent-memory systems?

The memory experiments use only 30–53 initially correct pairs per condition. How sensitive are the reported separation rates to the choice of histories, and can you provide results over more independently trained checkpoints?

Since readout closure is characterized using a classical congruence condition, what is the main theoretical insight beyond applying this established result to learned state representations?

1 comment

Thank you for asking about occurrence and consequences. We added sampled grid trajectories and replay tests on recorded tasks, and now state the selection rules and comparison cohorts in the revision.

Occurrence under sampled actions. We sampled 4,096 grid trajectories with uniform, distinct initial positions and object–direction actions, including blocked moves, and selected the first return involving movement within 32 actions. For two checkpoints, 258/4,096 and 644/4,096 had both current readouts correct and diverged at least once under the next 16 sampled actions. The earlier state's continuation is counterfactual (CPU float64, revised Table 1 and Appendix B.6).

Effects in recorded tasks. Revised §3 now states both denominators for the offline dialogue comparison. In 400 restaurant-search dialogues, we selected the first nonterminal repeat of the three annotated answers. For the selected adjacent turns where both sets of predictions were correct, replaying the same later utterances from the earlier state changed later slot answers in about 8–21% of cases across the three trackers, or about 3–8% of all dialogues. With a fixed catalog, gains in restaurant suitability outnumbered losses for each tracker. Rollback triggered by repeated predictions alone gave mixed results (Appendix B.9).

History in recorded chess games. The revised Discussion points to a known history-sensitive interface. Across 2,048 games, reloading a position snapshot and discarding history changed Stockfish's depth-eight recommendation after one or four recorded plies in 30 of 119 qualifying returns. Replaying moves from the snapshot after the most recent capture or pawn move restored the native recommendation at every tested nonterminal position (§6, Appendix B.10).

Future studies with access to deployment traces and resources for replay can use these tests to estimate how often such handoffs arise in ordinary operation and when carrying more history improves downstream outcomes.

Independent initializations and recurrent interfaces. We independently trained three additional grid models on the original data and tested 256 fresh pairs per history family. Two diverged and benefited from reconstruction in both families. The third was correct at all tested endpoints. In a separately trained memory model, 53/512 delete pairs and 63/512 overwrite pairs started correct on both branches. Of those, 47/53 and 52/63 diverged within 16 steps, respectively. Few were correct on both branches at all three 16-step endpoints. In the separate Qwen check, the first selected low-margin initially correct pair split after one common operation for each of two adapters sharing a backbone. All four fixed-selection pilot continuations stayed correct (§3, revised Tables 4–5 and Appendices B.2–B.3 and B.5).

Significance and theoretical scope. Classical congruence specifies which updates preserve current agreement over every continuation (original §2.1, Proposition 1). With a full-row-rank readout CzCz, affine updates Aaz+baA_a z+b_a, and columns of NN spanning ker⁡C\ker C, the requirement is CAaN=0CA_aN=0 for every action (original §2.1 and Appendix A). We apply this criterion to equal and correct readouts of the same sufficient source state, testing both continuation agreement and source correctness. A single reconstruction scored with the same updater exposed shared errors as well as repairing divergent predictions (original Tables 3–5, revised Figure 3 and Appendix B.7).

Presentation and prior work. The revised Introduction defines state and readout, and §5 relates our diagnostic to earlier history-pair and re-encoding tests. Both versions disclose AI assistance, and the authors take responsibility for the work.

Comments on the paper

1 comment

Two histories can give the same correct answers about the same sufficient source state and still predict different futures under identical actions. We use classical congruence to specify when current agreement should survive further updates (§§1–2).

We thank the reviewers for asking how the two-history tests relate to the guarantee for stores of named records. The revised abstract and Introduction identify a handoff failure.

In a store whose records can all be read out, we can rename identifiers or reorder slots and change the actions to match. After undoing that change, the predictions should agree. Revised §4.2 distinguishes two comparisons. The original slot-augmented SIT control tests whether enforcing state/action correspondence can retain predictive accuracy. The direct comparison with Structured and Canonical SIT tests whether the same scoring protocol distinguishes accuracy when the models are equally consistent under those changes. The nominal replacement transducer (NRT) demonstrates the formal correspondence by construction for finite legal rollouts in deterministic, exact arithmetic (Theorem 1). Structured SIT gives the strongest accuracy of the compared designs. NRT approaches Structured SIT in mean query accuracy across all four conditions and in full-state accuracy for tiered memory with 1,024 training transitions (original Table 7, revised Appendix E.5, Table 10).

Revised Figure 3 clarifies the original grid test. Rebuilding each state once from its own predicted positions raised the number of pairs correct on both branches at all three 16-step endpoints from 197/256 to 254/256, with the same updater. A separate condition showed why source scoring matters. Reconstruction corrected errors the original branches had shared (original Tables 3–4). Section 3 adds sampled returns and independently trained models to test whether the findings depend on selected histories or initializations. The revised Discussion uses recorded dialogue and chess to show how partial readouts can benefit from retaining the relevant history (§§3 and 6, Appendices B.9–B.10).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.