{"id":"1926fc30-e2b6-4cfd-8c87-86f51f981f29","arxiv_id":"2507.22281","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CoEx co-evolves an LLM agent's subgoal planning with an explicit, neurosymbolic belief state that is updated after each subgoal attempt, leading to improved success rates on ALFWorld, PDDL, and Jericho compared to several baselines.","lead":"This paper introduces CoEx, an LLM agent architecture that combines subgoal-level planning with a dynamically updated neurosymbolic world model, and tests it on three text-based agent benchmarks. It reports that CoEx outperforms several existing agent designs on planning and exploration tasks, though the evidence includes some inconsistencies and missing experimental controls.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim 'outperforms existing agent paradigms' is contradicted by the paper's own Table 4: WALL-E's reported total (95.00%) exceeds CoEx's (93.28%), yet §5.2 calls CoEx the highest; the quantitative basis for the headline needs correction.","rationale":"I read the paper's central claim as the empirical assertion that CoEx outperforms existing LLM agent paradigms in planning and exploration. The reader's weakest_assumption focuses on hallucinated belief-state updates, an important internal-validity concern for the mechanism. However, the most directly load-bearing problem is that the paper's own headline result in Table 4 contradicts the claim: WALL-E's reported ALFWorld total of 95.00% is higher than CoEx's 93.28%, while §5.2 states CoEx achieves the highest total success rate. Since WALL-E is presented as a baseline paradigm in §2 and Table 1, the abstract's 'outperforms existing agent paradigms' is unsupported as written. This does not require speculation about whether the belief state is accurate; it is a mismatch between the central claim and the reported evidence. The paper deserves credit for a detailed architecture description and a principled ablation against HiAgent, and the belief-state reliability problem is acknowledged in §5.5 rather than hidden. But the empirical headline must be corrected or statistically substantiated before the paper can be accepted. I therefore recommend a conditional verdict requiring revision of the performance claims and a reproducible recomputation of the ALFWorld aggregate, rather than outright rejection, because the broader comparison on PDDL and Jericho still shows improvement over the baselines tested there.","tokens_in":16560,"tokens_out":4279,"duration_ms":51440,"concrete_test":"Obtain per-task episode counts for the ALFWorld unseen test split (the paper does not report them) and independently recompute the overall success rate for WALL-E and CoEx from the per-task rows in Table 4. If WALL-E's total remains above CoEx's, revise the abstract and §5.2 to say CoEx is competitive with, rather than superior to, state-of-the-art baselines, or explicitly justify excluding WALL-E from the comparison. Additionally, report per-task N and confidence intervals; if the corrected totals tie or flip, the superiority claim requires statistical support rather than a point estimate.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract and §5.2 assert that CoEx 'outperforms existing agent paradigms in planning and exploration,' with §5.2 specifically claiming the 'highest total success rate' on ALFWorld. Table 4 reports WALL-E at 95.00% total versus CoEx at 93.28%, with WALL-E also ahead on Clean (97.00% vs 83.87%) and Heat (100% vs 91.3%). This is an internal contradiction, not a matter of missing error bars. The text later says CoEx achieves 'competitive performance with WALL-E,' which is accurate, but that does not reconcile with 'highest total success rate' or with the abstract's unqualified claim. Because WALL-E is explicitly categorized as an existing LLM agent paradigm in §2, the headline claim fails under the paper's own data. Moreover, the 'existing paradigms' claim is not broadly tested: PDDL (Table 5) compares only ReAct and HiAgent, and Jericho (Table 6) only ReAct. Thus the ALFWorld aggregate is the main quantitative support for the central claim, and that aggregate is not favorable. The load-bearing quantity is therefore the reported ALFWorld success rate, independent of whether the LLM verification and synthesis module is fully reliable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoEx, a hierarchical LLM agent architecture in which a subgoal-level Planner generates dynamic subgoal plans, an Actor executes them through action-level reasoning, and a neurosymbolic belief state (comprising code-based symbolic memory and LLM-synthesized textual memory) is updated after each subgoal attempt. The design aims to mitigate what the authors call exploitation bias and limited adaptation in monolithic LLM agents. CoEx is evaluated on ALFWorld, PDDL (Gripper and Blocksworld), and Jericho, benchmarked against ReAct, Reflexion, AdaPlanner, ExpeL, WALL-E, and HiAgent. The core claim is that CoEx 'outperforms existing agent paradigms in planning and exploration.'","tokens_in":16805,"tokens_out":3698,"duration_ms":41669,"significance":"The paper proposes a plausible and cleanly articulated architecture: separating subgoal-level planning from action-level execution, and making the world model explicit and updatable through a two-stage verification-and-synthesis process. The design is concrete, with pseudocode, prompts, and an example trajectory, and the cost analysis in RQ2 is a useful practical addition. If the empirical claims were fully supported, this would be a valuable contribution to LLM-agent research. However, the central quantitative claim is not supported by the paper's own data: the ALFWorld aggregate in Table 4 shows WALL-E outperforming CoEx, and the comparisons on PDDL and Jericho are much narrower than the abstract's unqualified claim. These issues must be resolved before the contribution can be assessed fairly.","major_comments":[{"comment":"The text states that CoEx 'achiev[es] the highest total success rate of 93.28%' on ALFWorld, but Table 4 reports WALL-E at 95.00% total, and WALL-E is also higher on Clean (97.00% vs 83.87%) and Heat (100% vs 91.3%). Since §2 explicitly categorizes WALL-E as an existing LLM agent paradigm, the abstract and conclusion claim that CoEx 'outperforms existing agent paradigms' is contradicted by the paper's own reported numbers. The later statement that CoEx 'achieves competitive performance with WALL-E' is accurate, but it does not reconcile with the 'highest total success rate' phrasing. This is a load-bearing inconsistency in the paper's headline claim and must be corrected.","section":"§5.2, Table 4"},{"comment":"The claim of outperforming 'existing agent paradigms' is supported on PDDL by comparisons only against ReAct and HiAgent, and on Jericho only against ReAct. The other baselines discussed in §2 and evaluated on ALFWorld—WALL-E, Reflexion, AdaPlanner, and ExpeL—are not run on these benchmarks. The cross-benchmark generalization of the headline claim therefore rests on a single task suite (ALFWorld), where CoEx is not the top performer. Please either add the missing baselines or explicitly qualify the claim as applying only to the methods actually compared in each setting.","section":"§5.2, Tables 5 and 6"},{"comment":"The paper acknowledges in §5.5 that 'LLM-based QA may introduce marginal inaccuracies during world model updates,' but it does not measure how often the Verification and Synthesis module (Stages 1 and 2) produces incorrect or hallucinated belief-state updates, nor whether such errors compound over the subgoal sequence. Because the Planner conditions all future subgoals on this belief state, reliability of the update mechanism is load-bearing. Please add a quantitative or qualitative error analysis of belief updates (e.g., a human evaluation of a sample of verification outputs and learned facts, or a measurement of error propagation).","section":"§4.4 and §5.5"},{"comment":"All results are reported as single-point success rates without confidence intervals, error bars, or significance tests, despite the substantial stochasticity of sampling from GPT-4o-mini across the Planner, Actor, and verification modules. The gap between CoEx (93.28%) and WALL-E (95.00%) on ALFWorld is small, and without multiple runs or a significance test, the ordering of the two methods cannot be treated as reliable. At minimum, the paper should report the number of runs or seeds used and discuss the sensitivity of the results.","section":"§5.1 and §5.2"}],"minor_comments":[{"comment":"There are several typos: 'effictiveness' in the introduction, 'myoptic' and 'W ALL-E' in §2, 'representave' in §5.1, 'maximimum' in §5.1, 'is is' in §4.4, 'promisining' in the Limitations section, and 'AlfWorld' (inconsistent capitalization) in §2.","section":"Throughout"},{"comment":"The pseudocode is internally inconsistent: Algorithm 1 line 16 assigns the return of ExecuteSubgoalEpisode to (εk, mk+1), but Algorithm 2's ExecuteSubgoalEpisode returns only Hexec; additionally, Algorithm 1 line 17 calls v.BeliefUpdate while Algorithm 2 defines the update through separate Verification and Synthesis procedures. Please align the naming and return values so the pseudocode matches the prose.","section":"Algorithms 1 and 2"},{"comment":"The table formatting is inconsistent (e.g., '100.00 %' with a space, and varying decimal places). Also, 'Picktwo' should be consistently capitalized (the task is 'PickTwo' in the ALFWorld literature).","section":"Table 4"},{"comment":"The sentence 'using an example from theZork1 of the Jericho text-based game' is missing a space and is awkwardly phrased; please rephrase.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The internal contradiction between the ALFWorld headline claim and Table 4 is serious enough to require explicit correction before any acceptance decision. I also recommend that the editor ask the authors to clarify how many trials/seeds underlie the reported point estimates and to add a more complete baseline matrix. The paper's architectural idea is interesting, but the current reporting does not meet the standard needed to support the abstract's broad claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: CoEx is a sensible architecture for LLM agents that update an explicit world model online, but the paper's own Table 4 undercuts the headline claim of outperforming existing paradigms. The combination is genuinely new: subgoal-level planning with a neurosymbolic belief state — code-based object memory plus LLM-generated textual memory updated through verification/synthesis after each subgoal. That is different from HiAgent (subgoal planning without an explicit world model), AdaPlanner (replanning without persistent memory), and WALL-E (offline rule learning). The appendix is a real strength: full prompts, symbolic-memory code, and a worked trajectory showing the world model evolving across four subgoal attempts. That detail makes the method reproducible-friendly.\n\nThe main problem is the ALFWorld claim. Section 5.2 says CoEx achieves the highest total success rate of 93.28%, but Table 4 reports WALL-E at 95.00%. WALL-E is an existing agent paradigm by the paper's own taxonomy, so the abstract's unqualified 'outperforms existing agent paradigms' is not supported by the paper's data. The later sentence that CoEx is 'competitive' with WALL-E is accurate, but it doesn't fix the overstatement. I'd also want error bars or significance tests; per-task numbers swing widely (Clean 83.87% vs WALL-E 97%, Picktwo 88.24% vs WALL-E 100%). Baselines on PDDL and Jericho are thin: only ReAct and HiAgent on PDDL, only ReAct on Jericho. And it's unclear whether the baselines use the same GPT-4o-mini backbone; the paper explicitly says that model for CoEx, not for baselines.\n\nThe weakest load-bearing premise is reliability of the LLM verification/synthesis module. If it hallucinates a belief update, the planner conditions on a corrupted world state. The paper acknowledges this in Section 5.5 but doesn't measure error rates or whether errors compound. That is a design concern, not a fatal one, but it deserves an experiment or at least a sensitivity analysis.\n\nThese are addressable defects. The architecture is coherent, the writing is clear, and the appendix shows real engineering effort. I'd send this to a serious referee, but I'd require the authors to correct the ALFWorld summary, add error bars, broaden baseline coverage, and speak directly to the hallucination risk in belief updates.","headline":"CoEx is a sensible architecture for online world-model adaptation in LLM agents, but the paper's own Table 4 contradicts its headline claim of outperforming all existing paradigms.","tokens_in":17352,"tokens_out":2008,"would_cite":true,"duration_ms":22299,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CoEx claims that LLM agents plan better when the world model is an explicit neurosymbolic belief state that co-evolves with subgoal-level exploration, and reports gains over ReAct, AdaPlanner, and HiAgent on ALFWorld, PDDL, and Jericho.","keywords":["LLM agents","world model","subgoal planning","exploration","neurosymbolic belief state","verification and synthesis","ALFWorld","text-based games"],"falsifier":"Run CoEx on ALFWorld while logging every learned fact and status line that the verification stage writes, then compare each against the simulator's true state. If a large share of false updates still leads to task success, belief-state fidelity is not the mechanism; if false updates reliably precede failed replanning, the paper's mechanism is confirmed.","tokens_in":16325,"feed_emoji":"🤖","tokens_out":11585,"duration_ms":111343,"temperature":0.7,"pith_summary":"The paper proposes CoEx, an LLM agent architecture whose central move is to make the world model an explicit, editable belief state instead of an implicit function of the LLM's parameters or context. A subgoal-level Planner generates goals, including deliberate exploratory goals, and delegates execution to an Actor; after each subgoal, a verification-and-synthesis module distills the raw experience into a neurosymbolic belief state combining code-based symbolic memory with structured textual memory. The paper's claim is that this co-evolution keeps planning aligned with the true state of a partially observable environment and overcomes the exploitation bias and limited adaptation of monolithic agents like ReAct. CoEx reaches a 93.28% total success rate on ALFWorld, exceeding ReAct, Reflexion, and AdaPlanner, and it outperforms ReAct on PDDL and Jericho in both success and progress rates. A sympathetic reader would take this as evidence that online world-model learning and subgoal-level exploration reinforce each other rather than pulling the agent in opposite directions.","feed_headline":"Agent that co-evolves plan and world model beats ReAct, AdaPlanner","feed_subtitle":"CoEx tops 93% on ALFWorld by writing each subgoal's discoveries back into an explicit belief state.","key_machinery":"The load-bearing machinery is the neurosymbolic belief state together with the two-stage update that refreshes it. $m_k$ is implemented as a Python class whose methods parse observations into symbolic predicates such as agent location, held objects, and on/clear relations, giving exact low-level state tracking without calling the LLM. $\\ell_k$ is generated by an LLM in two stages: Verification answers consistency and outcome questions about the subgoal trace, and Synthesis writes a status line, a justification, and a list of learned facts. The Planner is then a state machine whose states are these belief states; the Actor executes each subgoal in a bounded reason-act loop and signals completion or requests replanning. Because the update fires once per subgoal rather than per action, the extra computation is small, with the paper estimating that 70% of token cost stays in the Actor.","core_discovery":"The core discovery is that an LLM agent can be organized as a closed loop in which planning and world-model updates happen at the same subgoal granularity. The belief state is written $b_k=(m_k,\\ell_k)$, where $m_k$ is a code-based, object-oriented symbolic memory updated deterministically from observations, and $\\ell_k$ is a structured textual memory produced by the LLM-based Verification and Synthesis module after each subgoal attempt. The Planner is an LLM prompted as a state machine over this belief state: it reads its history $H_k$ and the latest $b_k$, emits a subgoal, and can regenerate the whole plan when new facts arrive. The paper argues that this design lets exploration be expressed as ordinary subgoals, and lets the outcome of exploration be written back into a persistent world model, directly addressing the two failure modes it identifies in monolithic LLM agents. Its experimental claim is that this architecture outperforms existing agent paradigms on ALFWorld, PDDL, and Jericho.","pith_inferences":["Editorial inference: the paper's strongest falsifiable prediction is that belief-state fidelity is the causal driver; feeding the verification module ground-truth labels instead of its own readings should preserve or improve performance, while injecting a single false learned fact should measurably degrade planning.","Editorial inference: the symbolic memory's predicates on PDDL could be handed directly to a classical planner, so a natural extension is to test whether the co-evolution loop helps most when subgoal semantics are fuzzy, as in text games, or also when planning could be fully symbolic.","Editorial inference: the decomposition of task-level exemplars into a skill library is a hidden variable, and transferring CoEx to a new environment likely depends on the breadth of that library; an ablation varying the skill library size would separate the contribution of the belief state from the contribution of the Actor's prior knowledge.","Editorial inference: if the update cost is as low as reported, the architecture suggests a general recipe for long-horizon LLM agents, subgoal-level planning, explicit memory, and verification at update boundaries, that could extend to domains such as web navigation or embodied instruction following."],"forward_implications":["Online world-model learning becomes a drop-in planning component: an agent can explore, record what it learns, and replan on the same clock, without an offline training phase.","Exploratory behavior no longer has to be demonstrated at the action level; the Planner can issue an exploratory subgoal such as 'investigate the trap door' as a first-class step, which is how CoEx handles open-ended discovery in Jericho.","Explicit belief states make the agent's model inspectable and correctable: a wrong learned fact can in principle be overwritten by a later verification-and-synthesis update rather than persisting implicitly in context.","The same architecture transfers across environment formats, from text-based embodied tasks to classical symbolic planning to interactive fiction, suggesting the mechanism is about the planning-update loop rather than a single benchmark's quirks."],"supporting_citations":[{"why":"Defines ReAct, the canonical action-level monolithic agent that CoEx critiques and uses as its primary baseline on all three benchmarks.","marker":"Yao et al., 2023"},{"why":"Provides Reflexion, the episodic self-reflection baseline that represents low-frequency world-model adaptation in the comparison.","marker":"Shinn et al., 2023"},{"why":"Provides AdaPlanner, the dynamic-replanning baseline that CoEx surpasses on ALFWorld, including the harder picktwo tasks.","marker":"Sun et al., 2023"},{"why":"Provides HiAgent, the monolithic subgoal-planning baseline used in the hierarchical-architecture ablation on PDDL.","marker":"Hu et al., 2025"},{"why":"Provides WALL-E, the offline rule-learning baseline whose ALFWorld score CoEx approaches without an offline learning phase.","marker":"Zhou et al., 2024"},{"why":"Provides the ALFWorld benchmark and its unseen test split used for the main household-agent evaluation.","marker":"Shridhar et al., 2021"},{"why":"Provides the Jericho interactive-fiction suite used to test exploration and open-ended discovery.","marker":"Hausknecht et al., 2020"},{"why":"Defines the PDDL domain description language used for the Gripper and Blocksworld planning tasks.","marker":"McDermott et al., 1998"},{"why":"Supplies the AgentBoard task setup and the progress-rate metric used for PDDL and Jericho.","marker":"Ma et al., 2024"},{"why":"WorldCoder, the code-based world-model approach that motivates the symbolic-memory design and marks the contrast CoEx draws with pure code-generator agents.","marker":"Tang et al., 2024"}],"fun_headline_variants":["CoEx: co-evolving world model and planning beats ReAct, AdaPlanner","LLM agent updates world model with each subgoal to outperform baselines","CoEx: neurosymbolic belief memory co-evolves with LLM planning","Agent writes subgoal outcomes into belief state, outpacing static planners","CoEx: co-evolving plan and world model tops ALFWorld, PDDL, Jericho"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the language model that checks and writes the belief-state updates reports what actually happened during a subgoal; if it quietly invents facts about the world, the Planner builds on a corrupted model and the co-evolution story fails.","fun_headline_variants_meta":{"raw":{"variants":["CoEx: co-evolving world model and planning beats ReAct, AdaPlanner","LLM agent updates world model with each subgoal to outperform baselines","CoEx: neurosymbolic belief memory co-evolves with LLM planning","Agent writes subgoal outcomes into belief state, outpacing static planners","CoEx: co-evolving plan and world model tops ALFWorld, PDDL, Jericho"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000377,"raw_usage":{"total_tokens":2006,"prompt_tokens":942,"completion_tokens":1064,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":959}},"tokens_in":558,"tokens_out":1064,"duration_ms":12034,"temperature":1.0,"reasoning_tokens":959,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:51:19.392519+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CoEx on ALFWorld while logging every learned fact and status line that the verification stage writes, then compare each against the simulator's true state. If a large share of false updates still leads to task success, belief-state fidelity is not the mechanism; if false updates reliably precede failed replanning, the paper's mechanism is confirmed.","supporting_citations":[],"review_version":1}