Internal World Model (IWM)

• IWM is an internal representation that predicts environmental state transitions and supports decision-making with structured latent models.
• Research showcases IWMs instantiated in transformer modules, JEPA-based vision predictors, and neurosymbolic belief states in LLM agents.
• Empirical studies reveal that enhancing representational fidelity in IWMs can boost planning efficiency and modular reuse across sorting, vision, and robotics tasks.

Searching arXiv for papers on Internal World Models and closely related world-model formulations. Searching arXiv for "Internal World Model" across transformers, vision, agents, and robotics. An Internal World Model (IWM) is an internal representation or predictive mechanism that enables an agent to estimate the state of its environment, anticipate transitions under actions or transformations, and use those internal predictions for reasoning or control. Recent arXiv usage is deliberately broad rather than uniform. In a small transformer trained to sort by adjacent swaps, an IWM is a structured latent representation that encodes global token order and supports a simple decision rule for selecting swaps (Bhalla et al., 21 Oct 2025). In JEPA-based vision, an IWM is a latent-space predictor conditioned on transformation parameters, trained to predict how masking and photometric transformations act on image representations (Garrido et al., 2024). In LLM-agent systems, the same term can denote an explicit neurosymbolic belief state that is updated online during interaction (Kim et al., 29 Jul 2025), a textual foresight module internalized within a single autoregressive policy (Zhang et al., 25 Jun 2026), or even a reusable slice of pretrained transformer layers treated as a fixed internal reasoning substrate (Jadouli et al., 20 Apr 2025). Formal work in robotics gives a more abstract characterization: an internal model emerges when the robot’s internal transition structure becomes sufficient, and in some cases isomorphic, with respect to the environment it interacts with (Weinstein et al., 2024).

1. Conceptual scope and defining criteria

A recurring distinction across this literature is between mere input–output competence and a genuinely structured internal model. The sorting study makes this distinction explicit: a model that simply maps a permutation to a good swap is not yet said to have an IWM. Instead, the term is reserved for an internal representation that captures global properties of the environment state and supports a simple, algorithmic decision rule from that representation to action (Bhalla et al., 21 Oct 2025). That usage sharply separates representational quality from end-task accuracy.

The JEPA-based vision formulation adopts the same idea at a different level of abstraction. There, the world is the latent space of image representations together with a transformation family, and the IWM is the predictor $p_\phi$ that maps a corrupted source representation, transformation parameters $a_{x \rightarrow y}$, and mask positions to a target latent representation. The key property is not invariance, but equivariance under explicitly conditioned transformations: the model learns how the visual world changes in latent space rather than merely ignoring such changes (Garrido et al., 2024).

LLM-agent papers push the term toward explicit state representation. CoEx defines the agent’s world model as a belief state $b_k = (m_k, l_k)$, where $m_k$ is symbolic memory and $l_k$ is structured textual memory, updated after each subgoal attempt through verification and synthesis (Kim et al., 29 Jul 2025). WorldMind likewise treats the LLM as an implicit world model, then aligns it with an external World Knowledge Repository containing Process Experience and Goal Experience, so that internal simulation becomes physically grounded rather than merely semantically plausible (Ren et al., 19 Jan 2026). The paper on “Internalizing the Future” goes further by embedding the world model directly into the policy’s token stream: the same autoregressive model produces a prospective rollout and a plan-conditioned success estimate before outputting the next action (Zhang et al., 25 Jun 2026).

Taken together, these formulations suggest that the term Internal World Model is best understood as a family resemblance concept centered on three properties: internal state abstraction, predictive or counterfactual transition structure, and downstream use in planning, control, or interpretability. The implementation may be a latent circuit, a transformer predictor, a symbolic belief state, or a reusable pretrained module, but the common role is to support internally grounded anticipation rather than reactive pattern completion.

2. Minimal mechanistic and formal formulations

The paper “Higher Embedding Dimension Creates a Stronger World Model for a Simple Sorting Task” provides one of the clearest mechanistic instantiations of an IWM (Bhalla et al., 21 Oct 2025). The environment state is a permutation $\pi$ of tokens $\{1,\dots,\ell\}$, the action is an adjacent swap at index $i$, the goal is the sorted permutation, and the reward is $+1$ if the permutation becomes sorted after a swap and $-0.001$ otherwise. The model is intentionally minimal: learned token and positional embeddings, a single-head causal self-attention block, no MLP block, and separate linear actor and critic heads. Embedding dimension $a_{x \rightarrow y}$0 is varied from 2 to 128, and PPO is used with $a_{x \rightarrow y}$1, learning rate $a_{x \rightarrow y}$2, and clipping coefficient 0.1.

The central result is that task performance saturates early, but world-model quality continues to improve with width. For sequence length 6, average accuracy is near 100% already by embedding dimension around 6 with 10M steps, and 99.2% of agents achieve 100% accuracy when $a_{x \rightarrow y}$3. Yet the representational metrics continue improving until about $a_{x \rightarrow y}$4: global ordering fidelity in the last attention row reaches about 87% for length 6 and 78% for successful length-8 agents, while top-2 alignment with the difference-based decision rule exceeds 90% (Bhalla et al., 21 Oct 2025). The two recurrent mechanisms are especially concrete. First, the last row of the attention matrix $a_{x \rightarrow y}$5 monotonically encodes token rank. Second, the selected transposition tends to align with the largest adjacent difference

$a_{x \rightarrow y}$6

This makes the IWM both interpretable and operational: a one-dimensional ordering coordinate plus a deterministic swap rule.

A complementary formal treatment appears in “An Internal Model Principle For Robots” (Weinstein et al., 2024). There the question is not how to probe a neural circuit, but when an internal transition system can be said to mirror the environment. The central notion is sufficiency of an equivalence relation $a_{x \rightarrow y}$7 on a deterministic transition system: $a_{x \rightarrow y}$8 If this condition holds, equivalence classes are preserved under all actions, so the quotient system remains dynamically predictive. The paper proves that minimal sufficient refinements exist uniquely, and that under appropriate conditions the quotient of the robot’s internal history space by such a refinement becomes bisimulation equivalent, and in some cases isomorphic, to the external environment (Weinstein et al., 2024). In that formalism, an IWM is not a heuristic latent code but an internal transition structure guaranteed to preserve environmental dynamics.

These two lines of work occupy opposite ends of the same spectrum. The sorting paper shows a concrete, emergent circuit in a tiny trained transformer; the robotics paper characterizes, at the level of transition-system theory, when an internal structure deserves to be called a model of the world at all.

3. Latent predictive IWMs in vision and model-based control

In vision, the paper “Learning and Leveraging World Models in Visual Representation Learning” defines an Image World Model as a predictor network $a_{x \rightarrow y}$9 operating in latent space, conditioned on transformation parameters and mask positions (Garrido et al., 2024). The target view $b_k = (m_k, l_k)$0 is a rich augmented image; the source view $b_k = (m_k, l_k)$1 is a more destructively transformed version with masking, grayscale, blur, solarization, and color jitter. Student and EMA-teacher ViT-B/16 encoders produce latents $b_k = (m_k, l_k)$2 and $b_k = (m_k, l_k)$3, and the predictor is trained with the JEPA-style loss

$b_k = (m_k, l_k)$4

The paper identifies three necessary ingredients for a strong IWM: conditioning, prediction difficulty, and capacity. Without conditioning, the predictor cannot distinguish transformations and world-model performance collapses, with MRR approximately 0.00. With sequence conditioning MRR is about 0.82; with feature conditioning, about 0.79. Stronger transformations and deeper predictors are likewise necessary to maintain equivariance rather than collapsing toward invariance (Garrido et al., 2024).

This framework explicitly treats the abstraction level of representation as controllable. Weak IWMs yield more invariant, contrastive-like features; strong IWMs yield more equivariant features that retain richer low-level information. The trade-off is visible in downstream results. An equivariant IWM with an 18-layer, 384-d predictor attains 67.5% linear ImageNet accuracy, but 83.3% with a frozen encoder and tuned predictor, and 84.4% end-to-end, outperforming the invariant variant and prior self-supervised baselines in that setting (Garrido et al., 2024). Here the IWM is not just a pretraining loss; it is a reusable latent dynamics module.

The review “World Models for Cognitive Agents: Transforming Edge Intelligence in Future Networks” places this within a broader model-based RL vocabulary (Zhao et al., 31 May 2025). A world model is an internal representation that encodes the environment in compressed latent variables and models temporal evolution, typically through an encoder, latent transition model, and decoder or task head. Its concrete instantiation, Wireless Dreamer, trains a latent world model to predict future latent states and rewards, then learns a Q-network mainly from imagined trajectories. In the UAV trajectory-planning case study, reward prediction achieved a mean absolute error of 0.359, approximately 5% relative error, and Wireless Dreamer reached near-asymptotic performance with approximately 46.15% fewer episodes than DQN while achieving a higher final average reward, 923.55 versus 829.04 (Zhao et al., 31 May 2025). This usage returns to the classical world-model role: internal simulation for sample-efficient planning.

4. Agentic IWMs: explicit belief states, planning trees, and textual foresight

LLM-agent work has turned IWM into an architectural design problem. “Reasoning with LLM is Planning with World Model” formalizes reasoning as an MDP whose states are natural-language descriptions of what is currently true and whose actions are reasoning steps (Hao et al., 2023). The same LLM serves as both agent and world model, with transitions of the form $b_k = (m_k, l_k)$5, and MCTS provides strategic search. The gains are concrete. In Blocksworld, RAP with LLaMA-33B reached 0.88 success on 4-step planning and 0.42 on 6-step planning, compared with GPT-4 CoT at 0.63 and 0.40. On GSM8K, RAP$b_k = (m_k, l_k)$6 plus aggregation reached 51.6%, above CoT plus self-consistency at 46.8%. On PrOntoQA, RAP obtained 94.2% prediction accuracy and 78.8% proof accuracy (Hao et al., 2023). In this setting, the IWM is an internal simulator over symbolic reasoning states.

CoEx externalizes that simulator into an explicit belief state. Its planner operates over subgoals, its actor executes them at the action level, and its adaptive world model is the neurosymbolic state $b_k = (m_k, l_k)$7, where symbolic memory records object-level facts and structured textual memory records status, verification outcomes, and learned facts (Kim et al., 29 Jul 2025). That co-evolving belief state materially improves long-horizon exploration. CoEx reports 93.28% overall success on ALFWorld, 88.24% on the hard “picktwo” subset, 73.3% success and 92.8% progress on PDDL, and 25.0% success with 55.5% progress on Jericho, substantially above ReAct on those benchmarks (Kim et al., 29 Jul 2025).

A different internalization strategy appears in “Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning” (Zhang et al., 25 Jun 2026). Instead of a separate simulator or value head, a single autoregressive model is trained to emit a prospective state rollout and a plan-conditioned confidence before choosing an action. The foresight spans are explicitly tagged as <imaginary>, <keyword>, <analysis>, and <confidence>, making the IWM a textual analogue of a world model plus a $b_k = (m_k, l_k)$8-like estimate. The paper argues for a three-stage pipeline—World Model Agentic Mid-Training, Format-Eliciting SFT, and Foresight-Conditioned RL—because post-training alone leads to superficial format imitation without grounded predictive capability. The full pipeline yields a search average of 50.6, above the best state-only baseline at 48.7, and a math pass@30 of 60.0 with mean@30 of 29.5, above 56.7 and 27.7 for the best comparison system (Zhang et al., 25 Jun 2026). This suggests that an LLM’s IWM can be trained as an internal deliberative process rather than attached as an external planning module.

WorldMind complements this trend by aligning the implicit LLM world model with an explicit World Knowledge Repository built from prediction errors and successful trajectories (Ren et al., 19 Jan 2026). It augments embodied agents with Process Experience rules and Goal Experience heuristics, improving EB-ALFRED from 44.4% to 48.0% success for GPT-3.5-turbo and from 41.2% to 49.2% for GPT-4.1-mini; on EB-Habitat the gains are 43.6% to 48.8% and 41.6% to 50.8% (Ren et al., 19 Jan 2026). A plausible implication is that agentic IWMs are increasingly moving from implicit parametric priors toward hybrid parametric–symbolic systems that can be updated online.

5. Modular reuse: IWMs as transferable internal substrates

Some recent work treats an IWM less as a learned transition rule and more as a pretrained internal substrate that can be repurposed for new domains. “Deep Learning with Pretrained ‘Internal World’ Layers” uses layers 8–9 of Gemma‑3‑1B as a frozen internal-world module inside a wildfire predictor (Jadouli et al., 20 Apr 2025). Tabular wildfire features are projected into hidden size $b_k = (m_k, l_k)$9, passed through the frozen Gemma slice with 8 attention heads and 14.7M frozen parameters, and then mapped to a scalar fire probability by a small classifier. The full system has 37.7M parameters, but only approximately 5.0M are trainable. On the Moroccan wildfire dataset it achieved accuracy 0.8760, AUC 0.9344, precision 0.8313, recall 0.9433, and F1 0.8838, with recall emphasized as the highest among the conventional deep baselines (Jadouli et al., 20 Apr 2025). Here the IWM is operationalized as parametric memory in pretrained mid-layers.

The vision IWM paper reaches a related conclusion from a different direction. Its predictor is not discarded after pretraining but fine-tuned as a reusable head. For the equivariant IWM, frozen-encoder predictor tuning reaches 83.3% on ImageNet-1k, compared with 82.9% for encoder-only tuning, and end-to-end tuning reaches 84.4%. On ADE20K, predictor tuning reaches 46.8 and end-to-end tuning 47.0, again exceeding encoder fine-tuning for that model (Garrido et al., 2024). This reusability is central: the internal world model is valuable not only for learning representations, but also as a deployable downstream module.

These modular papers suggest a broadened sense of IWM. The world model need not always be a separately trained dynamics learner. It may also be a pretrained internal layer stack whose representational geometry is rich enough to function as a reusable reasoning or adaptation substrate once suitable interfaces are learned.

6. Evaluation, safety, and ongoing disputes

The recent literature is equally notable for how it evaluates and challenges IWMs. WR-Arena argues that standard evaluation overemphasizes next-state fidelity and undermeasures simulation capabilities required for intelligent behavior (Team et al., 26 Mar 2026). It defines three benchmark dimensions: Action Simulation Fidelity, Long-horizon Forecast, and Simulative Reasoning and Planning, all built around the transition view $m_k$0. The benchmark introduces metrics such as Multi-round Smoothness and $m_k$1 for consistency degradation. Across evaluated models, no system exceeded 65% on the long-horizon metrics; PAN was best overall, with 70.3 on agent simulation, 47.0 on environment simulation, 53.6 on transition smoothness, 64.1 on generation consistency, and improvements of +26.7% and +23.4% over a VLM-only planner on open-ended and structured planning respectively (Team et al., 26 Mar 2026). The implication is that many apparent world models remain weak precisely where internal simulation is supposed to matter most.

Safety analyses sharpen that point. “World Models: The Safety Perspective” argues that a world model is meant to predict future environmental states or fill in missing information so an agent can plan and behave safely, but current systems remain vulnerable to unreasonable scene generation, temporal inconsistencies, model exploitation, and miscalibration (Zeng et al., 2024). The survey emphasizes robustness, calibration, symbolic guardrails, and formal verification as missing components of trustworthy IWMs. This is not a peripheral concern: once a world model is placed inside a decision loop, its errors are no longer merely perceptual.

A separate line of critique asks whether current LLMs possess IWMs at all in a human-like sense. “LLM world models are mental” adapts pulley-system paradigms from cognitive science and finds mixed evidence (Robertson et al., 21 Jul 2025). In Study 1, the best model reached only 26.1% exact MA accuracy, though estimates correlated with ground-truth mechanical advantage and were strongly driven by pulley count. In Study 2, models distinguished a functional pulley system from a jumbled fake system with F1 $m_k$2. In Study 3, when asked to distinguish a functional system from a matched but non-force-transferring system, performance fell to F1 $m_k$3, effectively chance (Robertson et al., 21 Jul 2025). The authors conclude that LLMs may manipulate internal world models sufficient for coarse structural discrimination, but not for nuanced connectivity reasoning.

The paper “Internal World Models as Imagination Networks in Cognitive Agents” reaches a parallel conclusion from a very different methodology (Ranjan et al., 5 Oct 2025). Human imagination networks built from vividness questionnaires show coherent centrality correlations and clustering across populations, whereas LLM-derived networks often lack clustering and show lower correlations under different prompts and conversational-memory conditions. The paper interprets this as evidence that current LLM IWMs are not structurally similar to human imagination-based internal models (Ranjan et al., 5 Oct 2025).

The resulting debate is not whether IWMs exist in a single universal form, but which properties are necessary before the term becomes substantively explanatory. Current work supports at least four answers: structured and interpretable latent geometry, accurate transition prediction, reusable internal computation, and reliable support for planning under perturbation. No existing paradigm fully satisfies all four at once.