{"id":"0368378f-bdcc-4af7-aba9-8e5dba1f2d22","arxiv_id":"2608.02713","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An agent-centric reframing of world modeling, replacing physical state prediction with 'information transitions' organized into six proxy functions and three empowerment levels.","lead":"This paper proposes that AI agents should use 'world proxies' that return useful information, such as simulated outcomes, retrieved memories, or quality scores, instead of only predicting future physical states. The authors organize these proxies into six functional types and three levels of agent improvement, from choosing better actions to continuously co-evolving with the agent.","discovery_kind":"paradigm_shift","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unified Eq. 2 is a labeling, not a model, because no state-update or composition rule connects the six proxy types; the single information state space is asserted, not derived.","rationale":"The reader's weakest_assumption identifies the same soft spot: the single information state space in Eq. 2 is asserted rather than shown to support one transition map for all six proxy functions. My concern is slightly more specific: even the per-proxy formulas in Section 4 have incompatible output types, and no iteration or composition rule for s_{\\ell+1} is supplied, so Eq. 2 is currently satisfied by trivial mechanisms. That said, this is a position paper, and the taxonomy, tables, and honest open challenges carry real value; the weakness argues for the CONDITIONAL verdict the reader already chose, not for rejection. The paper's own Section 5 admits that trust, grounding, and evaluation remain open, which is consistent with treating Eq. 2 as an organizing label until a typed instantiation is provided. I recommend UNCHANGED, with the specific request that the authors either add an operational grounding condition or present one fully typed worked trajectory that makes Eq. 2 compositional.","tokens_in":33271,"tokens_out":4396,"duration_ms":49274,"concrete_test":"Take the L1 web-agent example in Section 3.1 and write the full trajectory under Eq. 2: specify a concrete S, the initial s_0, and the update that produces s_1 after the execution-proxy query 'click Purchase'. Then query the memory proxy and show s_2 is obtained from s_1 and u_mem in the same S. If the update rule is not defined, or if the type of S changes between steps (e.g., from environment state to agent context to retrieved memory), the paper should either supply the missing typed composition rule or weaken the claim that Eq. 2 is a unified model. A second check: show that the constant proxy WP(s,u) = 'ok' is excluded by the formal definition; if it is not excluded, add an operational groundedness/actionability constraint to Eq. 2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that dynamics, spatial, execution, memory, skill, and reward/verification proxies are instances of one information-transition mechanism—rests on Eq. 2 in Section 2.3. That equation is unconstrained: S is defined as 'may include physical state, observation, memory, knowledge, execution result, verification, guidance, and more', and WP is not given any state-update rule, consistency constraint, or composition principle. As written, any query-answer system satisfies Eq. 2: an LLM-as-judge, a hash table, or even a constant function that returns a fixed string is 'conditioned on an agent interaction' and returns 'information' in S if S is the set of strings. The three requirements in Section 1.3 (closed loop, grounding, actionable gain) are qualitative and not part of the formalism. The individual formulas in Section 4 also reveal type mismatches: the dynamics proxy outputs a state-reward pair (Section 4.2), the execution proxy outputs a state plus feedback (Section 4.4), and the reward proxy outputs a scalar judgment (Section 4.7). To call these one transition map, the paper needs a typed composition rule and an update rule for s_{\\ell+1}, but Sections 3.1-3.3 only patch agent state with \\oplus or Train/CoEvolve and never state how the proxy's internal state evolves. Without a rule for iterating the loop, L1-L3 are descriptions of use, not consequences of the formalism. Thus the claimed unification is currently a naming scheme; the taxonomy can still be useful, but the paper's central formal claim is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This position paper argues that world modeling for continually improving agents should be broadened from physical next-state prediction to agent-usable information transitions. It defines an Agent-Centric World Proxy by Eq. (2), organizes the design space into six proxy functions (dynamics, spatial, execution, memory/experience, skill, reward/verification) and three empowerment levels (L1 inference-time guidance, L2 training-time optimization, L3 agent-proxy co-evolution), and maps representative systems onto a functions-by-levels table. The paper closes with open challenges on fidelity, trust, safety, and evaluation. It is explicitly a roadmap rather than an experimental study; Table 6 is labeled illustrative.","tokens_in":33582,"tokens_out":4713,"duration_ms":47254,"significance":"The main contribution is a useful organizing framework: the six-function taxonomy and the L1-L3 ladder clarify how heterogeneous mechanisms (video prediction, NeRF rendering, code execution predictors, memory retrieval, skill libraries, reward models) can all serve as agent-facing feedback modules, and the functions-by-levels map highlights underexplored cells. The paper is honest about its status, explicitly marking Table 6 as illustrative and listing open challenges on fidelity, trust, safety, and evaluation. However, the claimed formal unification in Eq. (2) is currently a notational schema rather than a model; the paper's value rests on the taxonomy and agenda, not on the formalism. If the formal claims are tightened or reframed, the roadmap is a potentially influential contribution to the world-modeling and agent-learning communities.","major_comments":[{"comment":"The formal definition states ŝ_{ℓ+1} = WP(sℓ, u^F_ℓ) with S described as 'may include physical state, observation, memory, knowledge, execution result, verification, guidance, and more,' but imposes no update rule, no composition principle, and no consistency constraint on WP. As written, any query-answer system—an LLM-as-judge, a hash table, even a constant function returning a fixed string—satisfies Eq. (2) if S is taken to include the relevant output type. Since the abstract and Section 2.3 present this equation as the basis for treating all six proxy types as instances of a single information-transition mechanism, the formalism is load-bearing; the paper should either supply a typed state space with an explicit update/composition rule or explicitly reposition Eq. (2) as a conceptual schema rather than a formal model.","section":"§2.3, Eq. (2)"},{"comment":"The L1–L3 equations define how the agent side is modified (⊕, Train, CoEvolve), but they never specify how the proxy's internal state or sℓ evolves when the loop is iterated. Consequently the three levels are descriptions of how proxies are used in practice, not consequences of the formal definition; in particular, the CoEvolve(·) symbol in Section 3.3 is a placeholder for an unspecified joint-update process. The paper should either give a concrete iteration semantics for the closed loop or explicitly state that L1–L3 are empirical usage categories rather than formal consequences.","section":"§3.1–3.3"},{"comment":"The per-function formulas produce type-heterogeneous outputs: dynamics returns a state-reward pair (ŝ, r̂), execution returns a post-execution state plus feedback (ŝ_exec, ŷ_exec), and reward/verification returns a single evaluation value (v̂_eval). Without a typing or composition rule that embeds these outputs into a common information state space S, the claim that all six forms are instances of one transition map ŝ_{ℓ+1}=WP(sℓ,u^F_ℓ) is a labeling rather than a derivation.","section":"§4.2, §4.4, §4.7"}],"minor_comments":[{"comment":"The figures contain numerous OCR artifacts (for example, '%FTJHO4QBDF' and '\"HFOU' in Figure 1); the final version needs a pass to restore readable text in captions and inline labels.","section":"Figures 1–10"},{"comment":"The Reward / Verification row marks the L3 cell as 'emerging' but provides no representative citation; either add one or state explicitly that the cell is currently empty.","section":"Table 6"},{"comment":"The relationship between the interaction step ℓ and physical time t should be stated more carefully; Section 4.2 says the dynamics proxy reduces to classical prediction when ℓ aligns with t, but interactions can also be nested within a physical step, so the alignment condition is not just equality of indices.","section":"§2.3 and §4.2"},{"comment":"The world proxy symbol is rendered in inconsistent calligraphic forms across equations and text; unify the notation for WP.","section":"Notation"}],"recommendation":"major_revision","confidential_remarks":"The paper reuses the L1-L3 labels from the authors' related work [31], and a substantial share of the supporting examples and references come from the same group; this is not disqualifying, but the editor may wish to check novelty overlap with [31]. The main publication risk is the formal overclaim in Eq. (2), which is fixable by reframing the equation as a conceptual schema or by adding structural constraints; the taxonomic content itself is a solid basis for a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper gives the agents/world-model community a useful vocabulary—six proxy functions, three empowerment levels—but the formal equation it builds on is too loose to back the word \"paradigm shift.\" Read it for the taxonomy, not for the math.\n\nWhat's actually new: the synthesis of dynamics, spatial, execution, memory, skill, and reward/verification into one agent-centric \"information transition\" framing, and the L1-L3 empowerment scale explicitly separated from Chu et al.'s intrinsic-capability scale (Table 4). That's a genuine contribution. The worked example of a web agent moving from guidance to training signal to co-evolving partner is clear and instructive. The paper is also honest about its status: Table 6 is labeled illustrative, and the open challenges (trust, reward hacking, agent-centric evaluation) are real and well stated.\n\nThe soft spots, in proportion. The central formal claim is weaker than the rhetoric. Eq. 2—\\(\\hat{s}_{\\ell+1} = \\mathcal{WP}(s_\\ell, u_\\ell^\\mathcal{F})\\)—is satisfied by any query-answer function, since S is allowed to be almost anything and WP has no update rule or composition principle. The paper's qualitative criteria (groundedness, controllability, actionability) are sensible but not part of the formalism, so the unification of the six proxy types is a naming scheme, not a model. The stress-test note is right when it says the individual formulas in Section 4 have different output types (state+reward vs. state+feedback vs. scalar judgment) with no typed composition. For a position paper that is acceptable, but it means the contribution is the taxonomy, and the abstract's \"paradigm shift\" language overstates what is a reframing. The L1-L3 reuse from Chu et al. is acknowledged and explicitly differentiated, so that's handled well; self-citations are heavy but mostly on-topic.\n\nWho it's for: researchers in agents, world models, and RL who want a common vocabulary to separate memory, skill, verifier, and simulator work. It's a roadmap, not an empirical or theoretical result. I'd send it to peer review as a position paper, conditional on the authors softening the paradigm-shift claim and either operationalizing the definition or saying plainly that it is categorical, not formal. With that revision, it's a solid contribution to have on record.","headline":"Agent-centric world proxy taxonomy is a useful vocabulary for the agents/world-model community, but the formal unification in Eq. 2 is too loose to carry the paradigm-shift claim.","tokens_in":34175,"tokens_out":3443,"would_cite":true,"duration_ms":34893,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"World modeling for agents should be redefined as building environment-grounded world proxies that return agent-usable information transitions — execution results, retrieved memories or skills, verification signals — rather than only…","keywords":["world model","world proxy","agent-centric","information transition","continual improvement","feedback modalities","inference-time guidance","co-evolution"],"falsifier":"Build an agent-centric benchmark where the same world proxy is either available or withheld during an agent's planning and training, with all else fixed; if agents with proxy access do not show measurable gains in decision quality, learning speed, or final performance, the claim that world proxies provide actionable information gain is falsified. A more structural test: find two proxy functions, such as execution simulation and reward verification, that cannot be represented by a single shared information state without losing modality-specific structure, which would show the unifying equation is nominal.","tokens_in":33013,"feed_emoji":"🤖","tokens_out":7457,"duration_ms":64386,"temperature":0.7,"pith_summary":"Continually improving agents need feedback they can act on, but real-environment interaction is expensive, slow, unsafe, and hard to parallelize. This paper argues that world modeling, classically a predictor of the next physical state, should be recast as an Agent-Centric World Proxy: an environment-grounded mechanism that answers agent-initiated queries with information transitions, such as execution outcomes, retrieved experiences or skills, and reward or verification signals. It organizes the design space into six proxy functions (dynamics, spatial, execution, memory/experience, skill, reward/verification) and three levels of empowerment for the agent (inference-time guidance, training-time optimization, and agent-proxy co-evolution). The payoff, if the framing holds, is a common language and roadmap for building proxies that make agents plan better, learn faster, and improve continually, with evaluation shifted from visual realism to actionable information gain.","feed_headline":"World modeling shifts from next state to useful feedback","feed_subtitle":"Recasting world modeling as six agent-usable feedback proxies for planning, learning, and co-evolution.","key_machinery":"The load-bearing object is the formal proxy map $\\hat{s}_{\\ell+1} = \\mathcal{WP}(s_\\ell, u^\\mathcal{F}_\\ell)$ together with its two generalizations: the interaction step $\\ell$ replaces the physical time step $t$, and the information state space $\\mathcal{S}$ replaces the physical state. The proxy function $\\mathcal{F}$ indexes six concrete instantiations — dynamics, spatial, execution, memory/experience, skill, and reward/verification — each answering a different question for the agent but all expressible in the same transition form. This single map is what lets the paper treat retrieval, simulation, and evaluation as one unified mechanism, and it supplies the \"information gain\" criterion that separates a useful proxy from a mere simulator or database.","core_discovery":"The paper's central claim is that the unit of world modeling should change from a physical state transition to an agent-usable information transition. Where a classical world model predicts $\\hat{s}_{t+1} = \\mathcal{WM}(s_t, a_t)$, an Agent-Centric World Proxy is defined as $\\hat{s}_{\\ell+1} = \\mathcal{WP}(s_\\ell, u^\\mathcal{F}_\\ell)$, with $\\ell$ an interaction step (not necessarily physical time), $s_\\ell$ an information state in a space that \"may include physical state, observation, memory, knowledge, execution result, verification, guidance, and more\", and $u^\\mathcal{F}_\\ell$ an agent-initiated query under proxy function $\\mathcal{F}$. Under this definition, a memory lookup, a code execution, and a reward estimate become moves of the same game: each delivers information the agent did not possess an instant earlier. The paper claims this shift broadens world modeling from future-state prediction to six functional forms and shows how they empower agents at three progressive levels, culminating in a closed loop where real-environment evidence updates both proxy and agent.","pith_inferences":["If the unified formalism is to be more than a label, a natural stress test is to train a single transition operator that jointly serves two structurally distant proxy functions, say execution simulation and reward verification, and see whether one shared information state can carry both.","The L1-L3 ladder suggests a concrete deployment recipe: ship cheap reversible inference-time guidance first, upgrade to training-signal generation once trust is established, then close the co-evolution loop — a sequencing the paper motivates but does not prescribe.","The framework implies a specific benchmark design: hold the proxy fixed, vary access to its feedback, and measure downstream agent performance; this \"information gain\" metric is the natural operationalization of the paper's core criterion and would make different proxy functions comparable."],"forward_implications":["Evaluation of world models should measure agent-centric information gain — whether feedback improved planning, learning, or verification — rather than realism or prediction accuracy alone.","Research areas now studied separately (video prediction, novel-view synthesis, browser and GUI simulation, memory retrieval, skill libraries, reward modeling and verification) can be described as instances of one design space and compared across proxy functions and levels.","A single proxy function can deepen its role over time: the paper walks one execution proxy from L.1 advisor that imagines a webpage before a click, to L.2 teacher that scores simulated rollouts into training data, to L.3 partner that is retrained from the agent's real deployments.","Sparse cells of the functions-by-levels map, notably spatial and reward/verification proxies at the co-evolution level, are flagged as largely open territory for future work.","The L3 loop, where real-environment evidence updates the proxy and distilled proxy knowledge updates the agent, is presented as both a learning mechanism and a safety mechanism that keeps proxy and policy aligned."],"supporting_citations":[{"why":"Supplies the classical world-model definition as a state transition model that the paper generalizes.","marker":"[57]"},{"why":"Exemplifies learning behaviors by latent imagination, the model-based RL lineage the dynamics proxy builds on.","marker":"[59]"},{"why":"Establishes the planning and learning loop through an intermediate model that the world proxy extends.","marker":"[146]"},{"why":"Provides the intrinsic L1-L3 capability scale of world models whose shorthand the paper deliberately reuses for its agent-centric axis.","marker":"[31]"},{"why":"Supplies the worked example of an execution proxy at L1: model-based planning for web agents that imagines page states before acting.","marker":"[52]"},{"why":"Supplies the memory and experience feedback mechanism used as a canonical instance of a memory proxy.","marker":"[140]"},{"why":"Supplies the skill-library and internalization mechanism used for the skill proxy and proxy-to-agent distillation.","marker":"[157]"},{"why":"Supplies the L3 co-evolution instantiation: a web agent whose world model is retrained from real trajectories in tandem with the policy.","marker":"[41]"},{"why":"Supplies the preference-optimization objective through which L2 proxy feedback is converted into policy training signals.","marker":"[128]"}],"fun_headline_variants":["World modeling: from state prediction to agent feedback","Redefining world models as agent-usable feedback","Six proxy forms for agentic world modeling","Agent-centric world proxies replace state transitions","World proxies: feedback for continual agent evolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes that a single information state space $\\mathcal{S}$ — meant to hold physical states, observations, memories, execution results, verification, and guidance — can coherently support one transition map $\\hat{s}_{\\ell+1} = \\mathcal{WP}(s_\\ell, u^\\mathcal{F}_\\ell)$ across all six proxy functions; if those feedback modalities do not share compatible structure, the unified formalism is a label rather than a model.","fun_headline_variants_meta":{"raw":{"variants":["World modeling: from state prediction to agent feedback","Redefining world models as agent-usable feedback","Six proxy forms for agentic world modeling","Agent-centric world proxies replace state transitions","World proxies: feedback for continual agent evolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000164,"raw_usage":{"total_tokens":1298,"prompt_tokens":1050,"completion_tokens":248,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":666,"completion_tokens_details":{"reasoning_tokens":180}},"tokens_in":666,"tokens_out":248,"duration_ms":3232,"temperature":1.0,"reasoning_tokens":180,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:09:13.851859+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build an agent-centric benchmark where the same world proxy is either available or withheld during an agent's planning and training, with all else fixed; if agents with proxy access do not show measurable gains in decision quality, learning speed, or final performance, the claim that world proxies provide actionable information gain is falsified. A more structural test: find two proxy functions, such as execution simulation and reward verification, that cannot be represented by a single shared information state without losing modality-specific structure, which would show the unifying equation is nominal.","supporting_citations":[],"review_version":1}