- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper investigates the relationship between recoverable world knowledge and behavioral competence in Transformer-based Sokoban agents. The authors train decoder-only Transformers to predict optimal actions from an initial board and action history, comparing eight training conditions in a factorial design that varies auxiliary supervision for state, action effects, and goal occupancy. They evaluate world-knowledge representations using linear, nonlinear, and attention-pooling probes, and assess behavioral competence through optimal-action prediction, matched-state decisions, and puzzle-solving rollouts under different trajectory conditions. The main finding is that explicitly supervising world knowledge substantially improves state recoverability without consistently improving solving performance from original states. Meanwhile, action-only models can distinguish states requiring different optimal actions despite incomplete state recovery, and explicit supervision provides benefits in some continuation settings involving non-optimal action histories. The authors also conduct activation interventions to investigate the behavioral relevance of decodable state representations. Overall, the paper highlights a substantial gap between explicitly recoverable world-state representations and downstream decision-making performance.
Critical strengths
1. Comprehensive experimental design and carefully controlled behavioral evaluations.
The paper presents a systematic (2^3) factorial comparison of state, action-effect, and goal-occupancy supervision, while keeping optimal-action prediction as the shared primary objective. This design allows the authors to examine how different components of world-knowledge supervision affect representation recoverability and downstream solving, including their combinations and interactions. The evaluation is also notably thorough. In particular, the matched-pair experiments control for initial boards, local observations, immediate action effects, and several action-history cues while requiring different optimal actions, providing strong evidence that the policies distinguish states beyond these local cues. The supplied-history experiments further distinguish on-plan states, detours with on-plan box configurations, and off-plan box configurations, enabling more targeted comparisons of state tracking and recovery after non-optimal histories. Together with the multiple random seeds, probe controls, and supplementary intervention analyses, these experiments provide a carefully controlled investigation of the relationship between world knowledge and behavior, while acknowledging the limits of causal attribution.
2. An informative empirical characterization of the gap between world-state recoverability and behavioral competence.
The paper investigates an important question concerning whether more explicitly recoverable world knowledge necessarily leads to better decision-making. Its results reveal a substantial dissociation, i.e. directly supervising the state representation increases complete-state recovery rate but does not improve original-start solving on the hardest puzzles. At the same time, action-only models can make accurate state-sensitive decisions despite relatively poor complete-state recovery, while explicit knowledge supervision improves continuation after certain supplied detour histories. These findings demonstrate that the relationship between recoverable world-state information and behavioral performance depends on the evaluation setting, rather than following a simple monotonic relationship. Although related distinctions between representation quality and downstream performance have been explored in earlier world-model research, this paper provides a particularly controlled analysis using exact environment states, solver-defined optimal actions, and component-specific supervision. The resulting observations are valuable for understanding both the capabilities and limitations of representation-based evaluations of learned world models.
Critical weaknesses
1. Insufficient mechanistic understanding of the dissociation between world-state recoverability and behavioral competence.
The paper provides extensive evidence that world-state recoverability and behavioral competence do not necessarily improve together, but offers limited mechanistic insight into why this dissociation occurs. Although the probing experiments identify decodable representations of player and box positions, action effects, and goal occupancy, the paper does not investigate how the Transformer computes and updates this information from action histories, or how it uses such information to select optimal actions. In particular, it remains unclear whether the model reconstructs the current state from the action history at each prediction position, reuses intermediate state representations through a state-transition-like computation, or relies on alternative task-relevant representations. Nor does the paper identify the attention heads, MLPs, or computational pathways responsible for state tracking and optimal-action selection. This limitation is particularly important given the paper's own findings: action-only models make accurate state-sensitive decisions despite incomplete state recovery, while Appendix F.2 shows that removing state-probe directions can nearly eliminate state decodability without comparably impairing solving performance in the action-only and all-component models. These results suggest that the policies may rely on representations or computational pathways not adequately characterized by the measured linear state codes, making a mechanistic investigation especially important.
The causal interventions also remain limited in scope. Although Appendix F examines state-probe direction ablations and selective box-code replacements, these experiments do not establish whether the policy actually uses the decoded world-state representations to compute optimal actions. Since Sokoban is a deterministic Markov environment in which the optimal action depends only on the current complete state and fixed board layout, a more direct test would be to intervene on the model's internal representation and examine whether its predictions change consistently with a counterfactual board state. For example, jointly replacing representations of player and box positions, action effects, and goal occupancy with those corresponding to another valid state requiring a different optimal action could test whether the policy follows the counterfactual state's optimal action. The existing box-code interventions do not establish such full-state causal consistency, leaving unresolved whether the recoverable world knowledge mediates the policy's decisions or whether alternative internal computations support its behavior.
2. Insufficient isolation of the functional effects of world-knowledge supervision.
Although the paper demonstrates that explicit world-knowledge supervision substantially improves complete-state recovery without consistently improving original-start solving, the experimental comparisons do not fully isolate the effects of the supervised information from changes in optimization and training resources. The conditions share the same training data, model architecture, and update counts, but the auxiliary objectives introduce additional computation and gradients. Moreover, the reduced advantage of all-component supervision under longer training suggests that differences in training efficiency may contribute to the observed behavioral effects. Without more closely matched computational budgets and auxiliary-objective controls, it remains unclear whether the differences between action-only and explicitly supervised models arise specifically from the additional world-knowledge targets or from more general changes in multitask optimization. This limits the interpretation of how explicit knowledge supervision affects behavioral competence.
Questions for authors
Clarification Requests
1. Interpretation of state-direction ablation results.
Appendix F.2 shows that removing state-probe directions nearly eliminates state decodability while leaving solving performance largely unchanged in the action-only and all-component models, but produces a larger behavioral effect under latent prediction and state matching. How do the authors interpret these differences across training objectives? Could they reflect different internal representations or computational pathways, redundancy in state information, or limitations of the identified linear state codes? Clarifying what these interventions establish about the causal role of world-state representations would help interpret the central recoverability–competence dissociation.
2. Role of optimization and training budget.
The training conditions match architecture, data, and update counts but not auxiliary computation or gradient contributions, while longer training substantially reduces the advantage of all-component supervision on supplied-history continuation. How much of the observed behavioral difference do the authors attribute to explicit world-knowledge supervision, as opposed to changes in optimization dynamics or training efficiency?
Suggestions for Future Work
1. Mechanistic analysis of state tracking and optimal-action computation.
Could the authors investigate how the Transformer internally computes and uses world-state information, particularly in light of the dissociation between state decodability and solving performance? For example, it would be valuable to identify attention heads, MLP components, or computational pathways involved in tracking player and box positions, modeling action effects, and selecting optimal actions. An important question is whether the model reconstructs the current state by aggregating information from the full action history at each prediction position, reuses intermediate state representations through a state-transition-like computation, or employs more compressed task-relevant representations without explicitly reconstructing the full board. Comparing these mechanisms across action-only and explicitly supervised models could clarify whether similar behavioral competence emerges from different internal computations.
2. Counterfactual world-state interventions.
Since Sokoban has deterministic Markov dynamics, could the authors conduct interventions that replace the model's internal representation of the current board with that of another valid state requiring a different optimal action, and test whether its predictions become consistent with the counterfactual state's optimal policy? Beyond the existing box-code replacements, jointly manipulating player and box positions, action effects, and goal occupancy, while controlling for unintended changes to other representations, would provide a more direct test of whether the decoded world-state information causally determines action selection. Comparing counterfactual optimal-action accuracy and downstream solving under these interventions would help distinguish functional state representations from merely decodable information.
3. Representation redundancy, recovery, and robustness.
Given the limited behavioral effects of state-direction ablation in some models, could the authors examine whether task-relevant state information remains accessible through other residual-stream directions, token positions, or layers after intervention? Tracking component-specific probe accuracy and behavioral performance in subsequent layers could reveal whether state information is reconstructed or whether the policy relies on alternative representations. Comparing these effects across action-only and all-component models would also help determine whether explicit supervision improves the robustness of world-state representations to targeted perturbations.
4. Compute- and optimization-matched controls.
Could the authors introduce compute-matched action-only baselines and additional auxiliary-objective controls with comparable loss scales or gradient statistics to distinguish the specific effects of explicit world-knowledge supervision from generic multitask optimization and additional training resources?
5. Generalization across task complexity.
Could the authors examine whether the dissociation between state recoverability and behavioral competence persists as the Sokoban environment becomes more complex, for example with larger boards, more boxes, or longer planning horizons? This would help determine whether explicit state representations become more functionally important as state-tracking and planning demands increase.
2 comments
Response to Reviewer pGTC (Part 2 of 2)
Part 2 of 2: optimisation, budget, Appendix F.2 and the suggestions
Weakness 2 and Clarification 2, optimisation and budget. On the 7x7 two-box board supervision raised complete-state recovery (all-component: 0.208 to 0.955), but our controls do not separate every optimisation effect from the supervised information. The cost of state supervision reported in the submission was largely a loss-balance effect: at coefficient 0.1, where gradient statistics are close to action only, state supervision solves +4.9 percentage points (pp) [+2.5, +7.4] more 32-60 puzzles (optimal length 32-60 moves) than at 1.0 (all three seeds; a descriptive comparison). Our rule for an improvement, fixed after the reviews and before the new runs and applied unchanged to the submission's runs, needs every seed positive and a mean of at least +2 points. No tested coefficient (0.03-1.0), site (block 8, 12 or 16) or budget (1x, 2x and 4x the 51,360 updates) meets it, and a small gain is not excluded: with six seeds at 0.1 the difference from action only is +1.6 pp [-0.2, +3.4] for state and +1.2 pp [-0.4, +2.9] for all-component supervision (Sections 4.3 and 6, Appendix K). Part of the advantage after supplied detours may be training efficiency: its point estimates are smaller at 2x and 4x (three seeds, Appendix H.4).
| After supplied detours, supervised minus action only, pp | 1x | 2x | 4x |
|---|---|---|---|
| All-component 1.0 | +4.7 | +0.7 | +0.6 |
| State 0.1 | +2.6 | +1.0 | +1.5 |
With six seeds at 4x, all-component 1.0 is -0.2 pp [-2.0, +1.6], and +2.3 pp [+0.4, +4.2] on a second set of such histories, the development panel's (Appendix H.4).
Clarification 1, Appendix F.2. Its "after" column read the removed probe on its own removed directions, which gives near zero by construction. Complete-state recovery with freshly fitted probes, three seeds:
| Objective | Linear, clean | Linear, after | MLP, clean | MLP, after | Linear, removed at every later block |
|---|---|---|---|---|---|
| Action only | 0.208 | 0.167 | 0.422 | 0.379 | 0.172 |
| Latent prediction + state matching | 0.317 | 0.236 | 0.619 | 0.534 | 0.238 |
| All-component | 0.955 | 0.913 | 0.957 | 0.932 | 0.913 |
The edit removes one linear code but not the state information, which fits your redundancy and linear-code alternatives; across objectives we can only describe the differences, with three seeds. After the edit, matched-neighborhood accuracy (both members optimal; pairs share the current neighborhood) is below the random-control range in every seed for latent prediction and all-component supervision, and within it for action only (revised Appendix F.2).
Suggestion 3. We tested this for the complete state only, not by token position: it stays recoverable after every later block's code is also removed (last column above), and supervision does not make behaviour more robust to the edit.
Suggestion 4. We did not match compute or gradients exactly. Besides coefficient 0.1 and action-only runs trained 2x and 4x as long (more compute than any supervised 1x run), we trained a control target of the same form: the state head at 0.1 tracking a different world, the real moves replayed under a fixed action permutation. It costs 32-60 solving, -5.7 pp [-8.3, -3.1] against action only (every seed), and the true-state target solves +7.2 pp [+4.7, +9.7] more: the outcome depends on the target, not only on adding an auxiliary loss. With one permutation at one site, this comparison does not isolate the information in the target from its optimisation effects, and its cost does not show that the true-state target helps (Appendix K).
Suggestion 5. At 51,360 updates, on 7x7 with three boxes recovery rises (0.047 to 0.921-0.944) and no supervised condition improves 32-60 solving, so the dissociation holds (state 1.0 is negative in every seed, mean -1.93 pp, under the 2 points a cost needs); on each two-box board from 8x8 to 11x11 at least two of the four supervised conditions improve it by the rule fixed in advance; at twice the budget the gain of all-component supervision at 0.1 persisted on 9x9 and was not shown to persist on 8x8 and 11x11, nor was that of state supervision at 0.1 on 8x8 and 9x9. We designed the 8x8, 10x10 and 11x11 boards and the budget controls after the 9x9 result was known and fixed their analysis before their runs. Lookahead difficulty, box density, prefix length and corpus change with the board, so we claim no trend or effect of board size and cannot say whether the benefit tracks planning demand (our comments to all reviewers; Appendices P and R).
Response to Reviewer pGTC (Part 1 of 2)
Part 1 of 2: the joint replacement and where the decision-relevant information is
Thank you for proposing the joint counterfactual replacement: it reveals a substantial but partial behavioural effect. In the all-component model, replacing the probe-defined code at every move position at block 8 (the supervised block) raises the rate of taking the donor's optimal action by 44 points, against 92 for the full residual swap and under 2 for random subspaces of the same rank. Its share of the full swap, 0.47 (seeds 0.46-0.48), is below the 0.5 that our criterion fixed in advance; the result shows neither full mediation nor necessity. A check added after these results removes four directions of the code that do not change the probes' outputs: the transfer is 36 points (share 0.39), still partial. After the edit the probes give the donor representation's predictions for the player and box cells at the current token in every case; this does not show that those predictions are correct. We also correct Appendix F.2, which you rightly questioned: freshly fitted probes still recover the state after its edit (complete state, action only: 0.167, unedited 0.208), so unchanged behaviour cannot show that the policy does not use state information.
Section numbers below refer to the revised PDF, which we have uploaded.
Weakness 1 and Suggestion 2, replacing the decoded state. On the paper's 7x7 two-box board (three seeds per objective, protocols fixed before the runs) we used 2,848 ordered pairs of histories that share the initial board, move count, player cell and local cues, differ in a remote box and have disjoint optimal actions. At each block from 8 to 16 we set the recipient's component in the span of the linear probes for player cell, box cells, goal occupancy and action effects (at most 64 directions) to the donor's, h' = h + UU^T(h_donor - h), where U is an orthonormal basis of the probe directions. Transfer is the rate at which the edited policy takes the donor's optimal action, minus that rate in the clean run. A code counts as carrying the decision-relevant state at a block only if, in every seed, its transfer is at least half that of the full residual swap and at least 10 points above the largest random subspace.
No block meets this in the absolute frame; in the action-only model the decoded code moves the decision by 6-8 points at blocks 8-11 (share at most 0.09; 2-3 points, the level of random subspaces, without the four directions). Replaced at the current token only, the code transfers within 0.5 points of random subspaces at blocks 8-14 in both models, and with same-state donors the current-token edit leaves the action unchanged in at least 95.9% of cases. With player-relative box directions in the action-only model (box positions are decoded more accurately in that frame), the result before the output block does not change. At the current token of block 16 the relative code met the criterion with the pre-specified code (60 points, share 0.68); without the four directions it transfers 0.2 points, like every code at the output block, so those transfers cannot be attributed to the decoded state alone, and without those directions no code meets the criterion at any block. No rollouts were run under the edit, so downstream solving under it is not measured.
Weakness 1 and Suggestion 1, where the decision-relevant information is. In full residual swaps on the same cases, swapping the current token alone first yields the donor's optimal action in at least half the cases at block 15, while swapping all other move positions still does so through block 14, in both models:
At block 8, swapping only the last eight positions gives 31.9% for action only (all positions: 93.7%) and 63.0% with all-component supervision: supervision makes recent positions more informative. Restricting the current token's attention to the board and its last 15 moves lowers the optimal-action rate from 94-96% to 78-80% in action only, latent prediction and all-component supervision, so in every model the current token draws directly on positions more than 15 moves back. These interventions show that the decision depends on representations at older history positions; they do not distinguish aggregating the history at each prediction position from reusing intermediate state computed at earlier positions, two of the alternatives in your review. A residual swap moves state and plan alike, so it locates this information without separating the two; none of these analyses identifies heads or MLPs or explains the dissociation itself (Appendix N).