acceptodds
Under review as a conference paper at ICLR 2027

Retaining History Is Not Maintaining State: StateCommit for Persistent World Models

Abstract

Persistent world modeling under partial observability requires maintaining interaction-established states that may remain relevant after they are no longer directly observable. Existing approaches commonly preserve past interaction evidence through long context, summaries, retrieval, or recurrent memory, and reconstruct the needed state when a query or prediction is made. However, retaining evidence does not guarantee a coherent representation of what is currently true, as accurate recovery of an individual state can coexist with substantial whole-state inconsistency. To this end, we introduce StateCommit, which shifts state resolution from read time to interaction time and incrementally maintains a query-independent Active State. Specifically, interaction evidence is converted into explicit state transactions that are verified before a deterministic runtime commits persistent updates. For visual prediction, slot-indexed Visual Bindings further ground the maintained semantic state in the prediction view induced by the known next action. We also introduce StateReturn, a multimodal benchmark for whole-state maintenance and state-consistent future prediction. On StateReturn, Full-History retains 98.61% target-state accuracy after 16 intervening interactions but only 13.89% full-state exact match, whereas StateCommit achieves 98.67–100.00% full-state exact match and reduces state-sensitive prediction error by 58.1% relative to the best memory baseline. A quick visual overview of this work is available at https://statecommit.github.io/.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Models, State Tracking, Partial Observability

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_AjT7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies persistent world modeling under partial observability: tracking the interaction-established environment states.

The authors argue that methods that retain interaction history (full context, summaries, recurrent memory) can answer single-slot queries accurately while failing to maintain a jointly consistent state.

To address this issues, the paper proposes StateCommit, which (1) maintains an explicit, validated and typed state table updated at interaction time; (2) for visual prediction, binds each slots to the frame observed at its last update.

They also introduce StateReturn, a benchmark that tests whether interaction-established state is maintained out of view and reflected in future predictions, which existing metrics (per-query accuracy, full-frame reconstruction) can miss. It contains AI2-THOR text episodes and MiniGrid visual worlds, with K unrelated interactions inserted before evaluation, and is scored by full-state exact match (FSEM) and counterfactual-paired pixel metrics. At K=16, Full-History reaches 98.61% target accuracy but only 13.89% FSEM, while StateCommit achieves 98.67–100% FSEM and reduces state-sensitive prediction error by 58.1%.

Critical strengths
  1. the paper identifies a clear and practically relevant failure mode: retaining full history can achieve high accuracy on single-slot query while fail to maintain whole-state consistencies, which indicates a gap in the per-query metrics.

  2. the evaluation design is detailed and thoughtful.

  3. the authors release the dataset and the implementation code, and the paper is clearly written with informative figures.

Critical weaknesses
  1. missing baselines. the core idea of StateCommit is not new: maintaining validated active records of relevant entities instead of reconstructing the state from full interaction history. several agent memory approaches, such as Mem0 and A-MEM, have adopted the same idea, maintaining a graph structure to store the facts about the entities. However the paper does not compare them with the proposed approach.

  2. Limited benchmark headroom. the proposed approach nearly saturates the StateReturn benchmark, which limits its value for future work.

Questions for authors
  1. Is there an ablation that removes runtime verification? According to Table 4, no proposal was ever rejected on Structured text or visual inputs, so verification cannot have affected those results. On Natural text, rejected proposals were all repaired. Without this ablation, the contribution of the transactional mechanism remains unclear.

  2. How do the missing baselines mentioned in the weaknesses section perform in the proposed benchmark?

6 comments
AuthorsOpenReview ↗

Response to Reviewer AjT7 (part 1)

We would like to thank the reviewer for the very insightful and valuable comments. Below are our responses to the comments in Weaknesses and Questions.

W1 & Q2: missing baselines. the core idea of StateCommit is not new: maintaining validated active records of relevant entities instead of reconstructing the state from full interaction history. several agent memory approaches, such as Mem0 and A-MEM, have adopted the same idea, maintaining a graph structure to store the facts about the entities. However the paper does not compare them with the proposed approach.

We thank the reviewer for highlighting these relevant approaches. We conduct additional comparisons with Mem0[1] and A-MEM[2], and clarify the distinction between existing agent memory and StateCommit below.

(1). StateCommit substantially outperforms agent-memory baselines in persistent-state maintenance. We evaluate Mem0 and A-MEM on StateReturn-Text (Natural). The results are presented below:

MethodAnswerEMTSEMFSEMTSATarget Coverage
Mem064.3934.616.5020.8538.44
A-MEM64.0046.6735.6765.6575.00
StateCommit98.6798.6798.6799.4898.67

StateCommit achieves 98.67% FSEM, outperforming A-MEM by 63.00 percentage points, with clear advantages in target accuracy and coverage.

(2). StateCommit differs from existing agent memory in how it maintains and uses world state. Mem0 incrementally updates entity-relation memories, while A-MEM organizes and evolves interconnected memory notes for retrieval. In contrast, StateCommit maintains a query-independent Active State through versioned transactions and atomic commits, enabling direct state readout and state-consistent visual prediction.

Reference:

[1]: Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory, ECAI 2025.

[2]: A-Mem: Agentic Memory for LLM Agents, NeurIPS 2025.

W2: Limited benchmark headroom. the proposed approach nearly saturates the StateReturn benchmark, which limits its value for future work.

We thank the reviewer for raising this important concern. We agree that StateReturn-Text leaves limited headroom. However, we would like to highlight two points below.

(1). StateReturn retains substantial headroom in State-Consistent Prediction. Although StateCommit achieves near-perfect FSEM in State Maintenance, its State-Consistent Prediction performance remains below saturation across interaction horizons. The results are summarized below:

MetricK=0K=1K=2K=4K=8K=16
SourceMAE ↓0.08320.12740.12230.12200.12480.1341
CFP ↑0.90280.86110.84720.81940.80560.7639

StateCommit achieves 83.33% CFP overall, decreasing to 76.39% at K=16. This indicates that translating accurately maintained states into consistent future predictions remains an open challenge within the existing benchmark.

(2). Imperfect visual evidence reveals further benchmark headroom. We further examine StateCommit under four visual evidence conditions, with results summarized below:

Visual EvidenceStateCommit TSEM (%)StateCommit FSEM (%)
Clean100.00100.00
Mild Sensor94.4471.53
Moderate Sensor49.3113.66
Partial Occlusion76.1644.91

StateCommit's FSEM drops from 100.00% under clean conditions to 13.66% under moderate sensor degradation. These results highlight opportunities to extend StateReturn toward more challenging visual evidence conditions.

AuthorsOpenReview ↗

Response to Reviewer AjT7 (part 2)

Q1: Is there an ablation that removes runtime verification? According to Table 4, no proposal was ever rejected on Structured text or visual inputs, so verification cannot have affected those results. On Natural text, rejected proposals were all repaired. Without this ablation, the contribution of the transactional mechanism remains unclear.

We thank the reviewer for this important question. We conduct both a Runtime-removal ablation and a controlled transaction fault-injection diagnostic to clarify the contribution of Runtime verification.

(1). Standard StateReturn evaluation does not fully expose Runtime's protective role. We replace the verified Runtime with an unverified Active-State updater and evaluate its effect on FSEM. The results are presented below:

Inputw/ Runtime FSEM (%)w/o Runtime FSEM (%)
Structured100.00100.00
Natural98.6798.67

Removing Runtime leaves FSEM unchanged for both Structured and Natural inputs, suggesting that the standard evaluation does not sufficiently stress the transaction failures targeted by Runtime verification.

(2). Runtime verification prevents state corruption under injected transaction faults. We conduct a paired diagnostic with 300 invalid transactions (100 each for stale versions, expected-value mismatches, and invalid typed updates) and 300 clean controls. The results are reported below:

Transaction TypeCountRuntime RejectionFSEM w/ RuntimeFSEM w/o Runtime
Stale Version100100.00100.000.00
Expected-Value Mismatch100100.00100.000.00
Invalid Typed Update100100.00100.000.00
Clean Control3000.00100.00100.00

Runtime rejects all 300 invalid transactions without rejecting any clean controls, preserving state integrity in every case. These results demonstrate the protection provided by deterministic transaction verification.

Thank you for the detailed response and the additional experiments. The new results on state-consistent prediction and degraded visual evidence address my concern about benchmark headroom, and the Mem0/A-MEM comparison and Runtime ablation are helpful. I have one follow-up question:

Q: StateCommit relies on a predefined state schema: a fixed set of attributes (place, openness, toggle state) with closed value domains. For general world models, deciding what state to track is a key part of the problem. How would the method extend to open-world settings where the relevant attributes are not known in advance? Also, were Mem0 and A-MEM given the same schema (e.g. in their extraction prompts to inform them what to track), or did they have to decide on their own what to store?

Reviewer_tPPpOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper distinguishes retaining the history of past interactions under partial observability from maintaining a coherent current state. It proposes StateCommit, which converts interaction evidence into explicit state transactions, verifies them, and incrementally commits them to a query-independent Active State at interaction time; for visual prediction, it aligns slot-indexed Visual Bindings to the prediction view before rendering. The authors further construct the StateReturn benchmark, which evaluates whole-state maintenance and state-consistent future prediction while controlling the number of intervening interactions, across text and visual tracks. They show that history-based methods retain target-state accuracy but lose full-state exact match, whereas StateCommit maintains nearly perfect FSEM and substantially reduces prediction error in state-sensitive regions.

Critical strengths
  • The problem formulation and evaluation design are clear. Evaluating target-state correctness (TSEM) separately from whole-state correctness (FSEM) addresses a point that is often overlooked in long-term memory evaluation. Through a shared protocol that controls the number of intervening interactions, a design that pairs factual and counterfactual futures, and metrics restricted to state-sensitive pixels, the benchmark isolates the effect of persistent state from other factors. Along with the dataset release, it provides a valuable evaluation framework.
  • The paper's structure and readability are good. By first presenting two research questions, then explaining the method and evaluation tasks in relation to them, and organizing the experimental results as numbered Observations, the paper makes it easy to follow the correspondence between claims and evidence. The figures also convey the key points accurately. The appendix provides detailed descriptions of the benchmark construction protocol, implementation, and hyperparameters, as well as additional analyses, including the experiment replacing the Grounder with a general-purpose VLM and the Binding Alignment diagnostic, reflecting a commitment to reproducibility and candid disclosure of limitations.
Critical weaknesses
  • In StateReturn, the tracked state components and their canonical value domains are predetermined. Furthermore, the intervening interactions are designed not to alter the source state being evaluated, although they may update other tracked slots. Under these conditions, once the interaction evidence is correctly interpreted and the corresponding transactions are committed, maintaining the Active State becomes a relatively straightforward deterministic process. This makes near-perfect FSEM unsurprising, particularly given that StateCommit's performance remains constant across K. The more challenging problem is correctly interpreting interaction evidence. However, the Grounder trained on the controlled MiniGrid setting achieves 100% EventFact exact accuracy on the validation set. According to the appendix, replacing it with the general-purpose Qwen2.5-VL-7B reduces FSEM to 9.72%. This substantial degradation suggests that the method's effectiveness depends heavily on reliable evidence interpretation. Evaluation in more realistic environments with imperfect or ambiguous evidence is needed to establish the robustness of the proposed approach.
  • On the text track, providing the complete interaction prefix to a strong model (the GPT-6 Astra Full-History control) yields FSEM of at least 99% for all values of K, nearly identical to that of StateCommit. This suggests that the reported failures of read-time state reconstruction may largely reflect the limited capability of Qwen2.5-Coder-7B-Instruct rather than a fundamental limitation of history-based reconstruction. Furthermore, in State-Consistent Prediction, StateCommit receives slot-indexed Visual Bindings aligned to the prediction view, whereas the history-based renderers receive compressed representations of past visual transitions without equivalent aligned visual inputs. Since Figure 3 shows that most of the improvement in SourceMAE and CFP comes from Binding Alignment, we cannot clearly attribute the gains to Active State maintenance. To isolate the contribution of explicit state maintenance, a comparison with a history-based baseline equipped with the same Visual Bindings, alignment procedure, and a capacity-matched rendering architecture is necessary.
Questions for authors
  • My understanding is that the proposed method emphasizes verification of state transactions before they are committed to the Active State. However, the contribution of each verification component remains unclear. Table 4 shows that the Reviewer is invoked only in Text-Natural, while Text-Structured and Visual State require no Reviewer calls. Moreover, the reported 100% repair rate refers to Runtime validity, not necessarily factual correctness. The Auditor is described as checking consistency with grounded evidence, but the paper does not specify its verification procedure or report rejection statistics. Since the Runtime only checks the Active State version, expected pre-update values, and typed value domains, it cannot detect factually incorrect but Runtime-valid transactions. Does the Auditor detect such errors, and how does it behave when the Grounder itself produces incorrect EventFacts? Could the authors report transaction-rejection statistics and ablations that remove the Reviewer and Auditor? These experiments would clarify whether the verification mechanism contributes meaningfully to performance beyond the accuracy of evidence interpretation.
5 comments
AuthorsOpenReview ↗

Response to Reviewer tPPp (part 1)

We sincerely thank the reviewer for the constructive comments. Below, we address the concerns raised in Weaknesses and Questions.

W1 (1): In StateReturn, the tracked state components and their canonical value domains are predetermined. Furthermore, the intervening interactions are designed not to alter the source state being evaluated, although they may update other tracked slots. Under these conditions, once the interaction evidence is correctly interpreted and the corresponding transactions are committed, maintaining the Active State becomes a relatively straightforward deterministic process. This makes near-perfect FSEM unsurprising, particularly given that StateCommit's performance remains constant across K.

We thank the reviewer for the insightful comment. We clarify the controlled evaluation design as follows.

(1). StateReturn is designed to isolate persistent-state maintenance from evidence interpretation. The predefined state domains and source-preserving intervening interactions provide a controlled setting for evaluating state persistence, while other tracked slots may still change.

(2). Source-preserving interactions do not make whole-state reconstruction trivial. Although the source state remains unchanged, intervening interactions may update other tracked slots. We summarize the Full-History results across representative interaction horizons below:

MetricK=0K=4K=8K=16
Full-History TSEM (%)100.00100.00100.0098.61
Full-History FSEM (%)100.0080.5630.5613.89

At K=16, Full-History achieves 98.61% TSEM but only 13.89% FSEM, highlighting the gap between target-state retrieval and coherent whole-state reconstruction.

W1 (2): The more challenging problem is correctly interpreting interaction evidence. However, the Grounder trained on the controlled MiniGrid setting achieves 100% EventFact exact accuracy on the validation set. According to the appendix, replacing it with the general-purpose Qwen2.5-VL-7B reduces FSEM to 9.72%. This substantial degradation suggests that the method's effectiveness depends heavily on reliable evidence interpretation. Evaluation in more realistic environments with imperfect or ambiguous evidence is needed to establish the robustness of the proposed approach.

We thank the reviewer for highlighting this important concern. We further evaluate StateCommit under imperfect visual evidence as follows.

(1). StateCommit maintains strong performance under mild visual degradation, despite imperfect grounding. To directly examine the grounding dependence, we evaluate four visual evidence conditions, including sensor degradation and partial occlusion. The results are presented below:

Visual EvidenceGrounder EventFact ExactMetricFull-HistoryStateCommit
Clean100.00TSEM99.54100.00
FSEM69.44100.00
Mild Sensor90.52TSEM85.1994.44
FSEM50.2371.53
Moderate Sensor48.45TSEM43.9849.31
FSEM12.6713.66
Partial Occlusion82.43TSEM75.0176.16
FSEM35.6544.91

Despite EventFact accuracy dropping to 90.52% under mild sensor degradation, StateCommit maintains 94.44% TSEM, with a 21.30-percentage-point FSEM advantage over Full-History. StateCommit also achieves higher FSEM under partial occlusion, while both methods experience performance declines under stronger sensor degradation.

AuthorsOpenReview ↗

Response to Reviewer tPPp (part 2)

W2 (Text): On the text track, providing the complete interaction prefix to a strong model (the GPT-6 Astra Full-History control) yields FSEM of at least 99% for all values of K, nearly identical to that of StateCommit. This suggests that the reported failures of read-time state reconstruction may largely reflect the limited capability of Qwen2.5-Coder-7B-Instruct rather than a fundamental limitation of history-based reconstruction.

We thank the reviewer for the insightful comments. We agree that sufficiently strong models can perform well. We further examine the accuracy and practical cost of different approaches as follows.

(1). Strong Full-History models achieve near-perfect accuracy, while smaller models show substantial performance gaps. We conduct additional evaluations with Qwen-VL-3B/7B and compare them with the GPT-6 Astra control. The results are presented below:

MethodStructured FSEMNatural FSEMTotal TimeAvg./QueryAPI Cost
Full-History (Qwen-VL-3B)63.1118.942h 35m 38s2.59sLocal
Full-History (Qwen-VL-7B)71.0645.394h 14m 59s4.25sLocal
Full-History (GPT-6 Astra)99.6799.174h 56m 34s4.94s$72.91
StateCommit (Qwen-Coder-7B)100.0098.673h 50m 16s3.84sLocal

(2). StateCommit achieves comparable accuracy with a locally deployable 7B model. Compared with GPT-6 Astra, StateCommit achieves near-identical FSEM with lower measured latency (3.84s vs. 4.94s per query), while avoiding the $72.91 API expense. This offers a practical alternative to repeated Full-History reconstruction.

W2 (Visual): Furthermore, in State-Consistent Prediction, StateCommit receives slot-indexed Visual Bindings aligned to the prediction view, whereas the history-based renderers receive compressed representations of past visual transitions without equivalent aligned visual inputs. Since Figure 3 shows that most of the improvement in SourceMAE and CFP comes from Binding Alignment, we cannot clearly attribute the gains to Active State maintenance. To isolate the contribution of explicit state maintenance, a comparison with a history-based baseline equipped with the same Visual Bindings, alignment procedure, and a capacity-matched rendering architecture is necessary.

We thank the reviewer for the valuable suggestion. We conduct an additional binding-matched control where both methods share identical Visual Bindings, differing in their access to the maintained Active State. The results are presented below:

KHistory SourceMAE ↓StateCommit SourceMAE ↓History CFP ↑StateCommit CFP ↑
00.09210.08320.87500.9028
10.17210.12740.69440.8611
20.16220.12230.69440.8472
40.16190.12200.65280.8194
80.16400.12480.68060.8056
160.17490.13410.63890.7639
Overall0.15450.11900.70600.8333

(1). Active State provides additional improvements beyond Visual Binding and Alignment. StateCommit improves overall CFP from 0.7060 to 0.8333 and reduces SourceMAE from 0.1545 to 0.1190, confirming the additional contribution of the Active State.

(2). The benefit of Active State becomes more pronounced at longer interaction horizons. At K=0K=0, both methods achieve relatively close performance. At K=16K=16, StateCommit improves CFP from 0.6389 to 0.7639 and reduces SourceMAE from 0.1749 to 0.1341, highlighting the value of persistent state information as intervening interactions accumulate.

AuthorsOpenReview ↗

Response to Reviewer tPPp (part 3)

Q1 (1): My understanding is that the proposed method emphasizes verification of state transactions before they are committed to the Active State. However, the contribution of each verification component remains unclear. Table 4 shows that the Reviewer is invoked only in Text-Natural, while Text-Structured and Visual State require no Reviewer calls. Moreover, the reported 100% repair rate refers to Runtime validity, not necessarily factual correctness.

We thank the reviewer for the insightful comments. We conducted additional Reviewer-removal ablations on Text-Natural as follows.

(1). The Reviewer substantially improves full-state correctness beyond Runtime validity.

MethodFSEMTSA
Full StateCommit98.6799.48
w/o Reviewer54.6784.51

Removing the Reviewer reduces FSEM by 44.00 percentage points, demonstrating its contribution beyond Runtime-valid repairs.

Q1 (2): The Auditor is described as checking consistency with grounded evidence, but the paper does not specify its verification procedure or report rejection statistics. Since the Runtime only checks the Active State version, expected pre-update values, and typed value domains, it cannot detect factually incorrect but Runtime-valid transactions. Does the Auditor detect such errors, and how does it behave when the Grounder itself produces incorrect EventFacts? Could the authors report transaction-rejection statistics and ablations that remove the Reviewer and Auditor? These experiments would clarify whether the verification mechanism contributes meaningfully to performance beyond the accuracy of evidence interpretation.

We thank the reviewer for this important question. We further examine the Auditor through ablations and transaction-level error detection experiments as follows.

(1). The Auditor's contribution is not fully reflected in standard-test ablations. We evaluate the effect of removing the Auditor on Text-Natural queries. The results are presented below:

MethodFSEMTSA
Full StateCommit98.6799.48
w/o Auditor98.6799.48

Removing the Auditor leaves FSEM unchanged on the standard test trace. We then conduct a targeted evaluation of its error-detection capability.

(2). The Auditor detects evidence-inconsistent transactions beyond Runtime checks. The Auditor verifies whether proposed state updates are supported by Grounder-provided EventFacts. To evaluate this capability, we construct 300 incorrect but Runtime-valid transactions, together with 300 clean controls. The results are shown below:

Transaction TypeAuditor RejectionsRejection Rate
Runtime-valid incorrect215 / 30071.67%
Clean control2 / 3000.67%

These results demonstrate the Auditor's ability to detect errors that pass Runtime validation.

(3). Runtime and Auditor provide complementary verification. Runtime enforces version, expected-value, and type constraints, while Auditor checks proposed updates against Grounder-provided EventFacts.

Reviewer_zb4AOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes that world models should maintain an explicit world state, which it calls "Active State", and use it to answer questions and predict future states. It introduces StateCommit, a system that updates the Active State through verified transactions derived from interaction evidence. The process has 4 stages: 1) extract facts from interaction evidence of a state change, e.g., door opening; 2) propose a typed update that specifies the expected old value and proposed new value; 3) deterministically check the proposal against the current state and allowed values; and (4) commit updates that pass verification. For visual prediction tasks, the system also stores a visual binding that associates a state value with a past view, and used it to render future predictions. Additionally, the paper introduces StateReturn, a benchmark that measures state accuracy and full-state consistency at varying intervening steps (K=0~16). The experiments show that StateCommit outperforms strong baselines, including full-history and memory-based approaches.

Critical strengths
  1. Explicit state representation and maintenance is an important and often overlooked aspect of world models. This is especially valuable for maintaining coherence over long horizons, where persisting individual facts (memory) are often insufficient.

  2. Treating state updates as database transactions with version checks and type constraints is an interesting and useful design choice.

  3. StateReturn provides a focused testbed for measuring state accuracy and consistency, and isolates the effect and noises of the agentic backbone.

Critical weaknesses
  1. Lack of downstream task success in an end-to-end agent setting.

  2. The approach relies on a slot-based state schema, which is environment-specific and has to be pre-defined, it can also be limited in its ability to represent relational and compositional facts.

  3. The paper could more clearly position StateCommit with contrast to prior approaches on state representation and tracking, such as LLM-based state tracking (Li et al., 2025; Rozanov & Rei, 2025), and agent memory systems that track evolving state (Fan et al., 2026). Clarifying how StateCommit differs from these approaches would make its contribution easier to assess.

  4. It remains unclear how the method scales to more complex scenarios or longer-horizon tasks with many state slots, such as computer-use or coding agents.

Refs:

  • Li et al., 2025: (How) Do Language Models Track State? ICML.
  • Rozanov & Rei, 2025: StateAct: Enhancing LLM Base Agents via Self-prompting and State-tracking
  • Fan et al., 2026: Can Agent Memory Systems Track Evolving State?
Questions for authors
  1. How does StateCommit perform on end-to-end task success rates when integrated into typical agentic pipelines, compared to full-history or memory-based baselines?

  2. How does the approach scale in terms of the schema size (i.e. number of slots)? Does its effectiveness hold as the state representation grows?

2 comments
AuthorsOpenReview ↗

Response to Reviewer zb4A (part 1)

We sincerely thank the reviewer for the constructive comments. Below, we address the concerns raised in the Weaknesses and Questions.

W1 & Q1: Lack of downstream task success in an end-to-end agent setting.

We thank the reviewer for the valuable comments. We conducted additional closed-loop agent evaluations to assess whether StateCommit improves downstream task success. We would like to explain it as follows.

(1). StateCommit supports successful downstream task execution through persistent state maintenance. We evaluate two tasks in MiniGrid: single-object pickup and ordered two-object manipulation, where the agent must relocate one object before retrieving another. The results are presented below:

MethodSingle-Object PickupTwo-Object Manipulation
Bounded-History25.004.08
Recurrent Memory66.6734.69
Full-History98.6165.31
StateCommit100.00100.00

(2). StateCommit achieves a larger advantage in tasks requiring multiple persistent states. Full-History drops from 98.61% to 65.31% on two-object manipulation, while StateCommit maintains 100% success on both tasks, yielding a 34.69-percentage-point advantage on the more challenging task. This highlights the downstream benefit of maintaining multiple persistent states for sequential task execution.

W2: The approach relies on a slot-based state schema, which is environment-specific and has to be pre-defined, it can also be limited in its ability to represent relational and compositional facts.

We thank the reviewer for the insightful comment. While StateCommit requires a predefined schema, it supports dynamically evolving world states and certain relational and compositional facts, as clarified below.

(1). The schema is predefined, but the world state is dynamically determined. StateCommit specifies the state components and their valid value domains, while the actual state values are established and continuously updated through interactions rather than predefined.

(2). Typed slots support relational and compositional facts. In StateReturn-Text, place(mug) = (inside, drawer) represents an entity relation, which can be combined with openness(drawer) = closed to express that the mug is inside a closed drawer. While the current representation supports such compositions, its expressiveness is limited to predefined relation types and does not cover arbitrary relational dependencies.

AuthorsOpenReview ↗

Response to Reviewer zb4A (part 2)

W3: The paper could more clearly position StateCommit with contrast to prior approaches on state representation and tracking, such as LLM-based state tracking (Li et al., 2025; Rozanov & Rei, 2025), and agent memory systems that track evolving state (Fan et al., 2026). Clarifying how StateCommit differs from these approaches would make its contribution easier to assess.

We thank the reviewer for highlighting these related studies. While they share the goal of tracking evolving states, StateCommit differs in its focus and mechanisms, as clarified below.

(1). StateCommit extends state tracking to whole-state consistency and state-consistent visual prediction. Beyond maintaining current states, StateCommit evaluates whole-state coherence and integrates semantic states with aligned slot-indexed visual bindings for future prediction.

(2). StateCommit maintains persistent world state separately from model-internal tracking or agent reasoning traces. Li et al.[1] investigate how Transformers internally track states, while StateAct[2] uses self-prompting and Chain-of-States to guide agent decisions. In contrast, StateCommit maintains a persistent, query-independent Active State through interaction-driven updates.

(3). StateCommit emphasizes guarded atomic transactions, complementing StateMem's supersession and dependency management. The recent StateMem[3] maintains evolving facts through structured state units, supersession, and dependency rechecking. StateCommit instead uses versioned transactions with expected-value checks and atomic commits to enforce structural consistency during state updates.

References:

[1]: (How) Do Language Models Track State? ICML 2025.

[2]: StateAct: Enhancing LLM Base Agents via Self-prompting and State-tracking.

[3]: Can Agent Memory Systems Track Evolving State?

W4 & Q2: Scalability to more complex and longer-horizon tasks with increasing numbers of state slots.

We thank the reviewer for the valuable comments. We conducted additional scalability experiments by independently varying the schema size and interaction horizon. We would like to explain it as follows.

(1). StateCommit maintains complete state accuracy as the schema size increases. We vary the number of state slots from S=3S=3 to S=48S=48, while fixing the interaction sequence length (T=96T=96) and query horizon (K=32K=32). The results are presented below:

MetricMethodS=3S=6S=12S=24S=48
TSEMFull-History10040245382
StateCommit100100100100100
FSEMFull-History00000
StateCommit100100100100100
Slot CoverageFull-History49.329.210.24.22.1
StateCommit100100100100100

StateCommit maintains 100% target-state accuracy, full-state accuracy, and slot coverage across all tested schema sizes. Full-History can recover individual target states but produces incomplete state maps.

(2). StateCommit maintains consistent state accuracy across longer interaction horizons. With S=24S=24 and T=96T=96 fixed, we increase KK from 0 to 64. The results are presented below:

MetricMethodK=0K=4K=8K=16K=32K=64
TSEMFull-History857361725369
StateCommit100100100100100100
FSEMFull-History000000
StateCommit100100100100100100

StateCommit maintains 100% target-state and full-state accuracy across all tested horizons. In contrast, Full-History achieves 53–85% target-state accuracy but 0% FSEM throughout, indicating its difficulty in reconstructing complete state maps despite recovering individual target states.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Dear AC and Reviewers,

Thank you for your valuable feedback. We have provided our responses and uploaded a revised manuscript, with all revisions highlighted in blue. We would be happy to address any remaining questions.

Best regards,

The Authors

Dear Reviewers,

The authors have submitted their rebuttal to address your concerns. As the discussion deadline approaches, please review their response and share any remaining concerns or questions soon so the authors have enough time to respond.

Thank you, -AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.