{"entries":[{"id":"arxiv:2610.26844","kind":"paper","title":"AnyGoal: On World Model Generalization Beyond Expert Rollouts","text":"The goal of a world model is to capture the dynamics of an environment. This has the great appeal that it allows agents to plan and simulate a course of actions with the help of a world model. For these uses to be reliable, however, the learned dynamics must behave well beyond the specific state-action combinations observed in expert rollouts. Despite this, the relationship between data coverage, dynamics generalization, and downstream planning performance remains underexplored. We study this relationship through the lens of compositional generalization: can a world model recombine familiar state and action factors into valid dynamics? To answer this, we introduce AnyGoal, a benchmark of goal-shifted task variants built from widely used world-model environments, together with an open-source library. Each task is designed to be within reach of a world model that has learned the underlying dynamics: individual state factors and actions remain within their training support, but their combinations are underrepresented in expert trajectories. Across multiple environments and world model families, we find that current models consistently exhibit failures under these recombinations. By examining dynamics and representation failures across model families, interaction regimes, and evaluation metrics, we study where current world models generalize and where challenges remain for goal-shifted prediction and planning. Together, this framework, benchmark, and analysis provide a basis for ","url":"https://arxiv.org/abs/2610.26844","source":"arxiv","published":null,"file":null,"sha256":null,"by":"worker:brain-x-snapshot","method":"snapshot","at":1792476044454,"changed":{"by":"owner","method":null,"at":1794680965863},"cites":[],"tags":["votes-hf-papers"],"public":true,"staleSince":null,"wordsAt":1792476044454,"snippet":" … This has the great appeal that it allows agents to [plan] and simulate a course of actions with the help of a [world] [model] … "},{"id":"arxiv:2607.01531","kind":"paper","title":"OPINE-World: Programmatic World Modeling with Ontology-error-Prioritized Interactive Exploration for ARC-AGI-3","text":"Learning a useful world model from minimal interaction is central to building agents that adapt to unfamiliar tasks. Programmatic world modeling approaches such as WorldCoder quickly learn transition models when given pre-supplied symbolic representations; however, such approaches are insufficient when the symbolic representation and goal are unknown and continually changing, and are intractable under large action spaces due to rigid planning heuristics. Extending the programmatic world modeling approach, OPINE-World additionally maintains an abstracted representation in which provisional objects, incomplete causal explanations, and unresolved interactions can exist in a Bayesian, exploration-centric hypothesis space. This enables learning an unknown, dynamically evolving state, transition, and goal. Exploration and program revision share this evolving abstraction, allowing new evidence to revise both the hypothesized mechanics and their implementation in tandem at test time. ARC-AGI-3 presents this open-world problem as a general-intelligence learning benchmark. Through our OPINE-world model learning paradigm, we improve the Relative Human Action Efficiency score of Claude Opus 4.8 high from 1.5% to 78.40% on the ARC-AGI-3 benchmark.\n\nMechanism: Hybrid world model kept per game: a qualitative object-oriented natural-language model (S_NL,T_NL,R_NL), a Bayesian belief state S_B of empirical object-type x action x context effect distributions built by a deterministic replay-based extractor, and executable Python transition/reward models plus planner and an observation-to-object abstraction. Every observation goes into a replay buffer D and updates S_B. Acceptance gate: the programmatic model must be replay-consistent (Phi_1): for all stored transitions, predicted successor fields and reward equal the observed ones. A mismatch stops the current plan, opens a short window collecting related counterexamples, and triggers synthesis repair (CEGIS-style). A consistent model plans toward reward (Phi_2: planned sequence reaches reward in the model, executed as an online test of the model); if no such plan exists, S_B posterior uncertainty directs epistemic experiments to find confounders. Only deterministic discrete envs are handled.\n\nEvidence: Claude Opus 4.8 high on 25 public ARC-AGI-3 games: RHAE 1.5% base to 78.40%; 20/25 games, 160/183 levels completed. Table 1 mean+-SE: Full 78.4+-6.3; No S_B 58.2+-8.3; baseline1 63.7+-7.4; DreamTeam 38.1+-7.6; Continual 20.5+-5.7; OpenClaw 5.2+-1.0 (games completed 20, 12, 14, 6, 3, 0 of 25). Removing S_B reduces RHAE by 20.24 points. Only system better on Community leaderboard as of July 25, 2026: NVIDIA NOOA on GPT-5.6-sol xhigh, 85.1%; paper notes a 4.6x RHAE gap between Opus 4.8 high and GPT 5.6 Sol xhigh base models. Per-game budget 2000 actions or stall of >500 actions per level.\n\nCost: A complete 25-game evaluation uses Claude Opus 4.8 at high reasoning effort; including replay and planning, individual environments can consume up to 2 billion cumulative tokens. No dollar or wall-clock figure.\n\nRead for the report on Memento 3 (https://cb4.endpointservices.workers.dev/library/research-memento-3): arXiv API record + arXiv HTML full text (read relevant sections). Could not verify: No code URL (supplementary material only). No per-task dollar cost. Not stated whether Table 1 is a single run per game; RQ2/RQ3 ablations use low effort and a single case game g50t. baseline1 in Table 1 scores 63.7 here vs 58.12 reported in 2605.05138 because paper says it uses matching game builds and scoring; model behind baseline1 in this table not verified. Limitations: only deterministic discrete environments; replay consistency gives no guarantee for unvisited objects. Author-flagged negative: in some games (e.g. tr87) Bayesian guidance investigates goal-irrelevant uncertainty. | code note: Source code is 'provided as supplementary material'; no URL given in the text.","url":"https://arxiv.org/abs/2607.01531","source":"arxiv","published":"2026-07-01","file":"corpus/research-memento-3/arxiv/2607.01531.pdf","sha256":"aa61cf48f9e7bdfae2f3d57568aed8d1aa51c5197dbaa2e57106491ba6c28a95","by":"owner","method":"research:memento-3","at":1791712838650,"changed":{"by":"owner","method":null,"at":1794680969746},"cites":[],"tags":["memento-3","agent-research"],"public":true,"staleSince":null,"wordsAt":1791713969439,"snippet":" … under large action spaces due to rigid [planning] heuristics. Extending the programmatic [world] [modeling] approach, OPINE-[World] additionally maintains an abstracted representation in which … "},{"id":"arxiv:2605.05138","kind":"paper","title":"Executable World Models for ARC-AGI-3 in the Era of Coding Agents","text":"We evaluate an initial coding-agent system for ARC-AGI-3 in which the agent maintains an executable Python world model, verifies it against previous observations, refactors it toward simpler abstractions as a practical proxy for an MDL-like simplicity bias, and plans through the model before acting. The system is intentionally direct: it uses a scripted controller, predefined world-model interfaces, verifier programs, and a plan executor, but no hand-coded game-specific logic. The agent-facing prompts, workspace, and controller contain no game-specific code, game-specific prompts, hand-coded heuristics, hidden solutions, or other game-specific information; the same agent and prompts are used across games. Because the coding agent has broad system access, we audit unintended information channels, describe earlier vulnerable harnesses, and explain how the current harness closes observed leakage channels while reducing benchmark-specific information exposure. We report results on the 25 public ARC-AGI-3 games. Each playthrough starts from a fresh agent instance and clean workspace, with no access to files or conversation state from earlier playthroughs. With GPT-5.5 high reasoning effort, the agent fully solved 15 games and achieved a mean per-game RHAE of 58.12%. With GPT-5.4 high reasoning effort, it fully solved 8 games and achieved a mean per-game RHAE of 41.29%. Performance on the private validation set, which is not yet available to us, remains to be tested. Overall, the results provide preliminary evidence that verifier-driven executable world models are a promising approach for ARC-AGI-3 agents. Full run artifacts are released with the code at https://github.com/astroseger/arc-3-agents-baseline1.\n\nMechanism: Baseline: a Codex CLI (v0.128.0) coding agent driven by a scripted external controller. The agent maintains a Python codebase as an executable world model (world_model_engine.py transitions, world_model_state_io.py state/rendering, world_model_main_planner.py planner, all starting as empty interface templates). A world-model verifier replays recorded observations; a planner verifier checks plans reach LEVEL_COMPLETED in the model; a plan executor simulates a plan in the model and runs it in the real game, stopping at the first predicted-vs-observed settled ASCII frame mismatch (treated as a blocking modeling error). Controller prompts repeated refactoring toward simpler abstractions as an MDL proxy, issues RESET on GAME_OVER, and restarts a stalled agent up to ten times. No game-specific code or prompts.\n\nEvidence: 25 public ARC-AGI-3 games, one fresh playthrough each. GPT-5.5 high reasoning: fully solved 15 games, mean per-game RHAE 58.12%. GPT-5.4 high reasoning: fully solved 8 games, mean per-game RHAE 41.29%. Variability: ft09 100.00% RHAE with GPT-5.4 vs 57.80% with GPT-5.5; wa30 49.77% (7/9 levels) vs 9.11% (4/9 levels) (which model is which on wa30: text says 'compared with ... GPT-5.5', so the 49.77% run is GPT-5.4).\n\nCost: A single ChatGPT Pro subscription (USD 200 per month) was sufficient for roughly two to eight games per weekly Codex limit with GPT-5.5 high. No token counts or per-game dollars.\n\nCode: https://github.com/astroseger/arc-3-agents-baseline1\n\nRead for the report on Memento 3 (https://cb4.endpointservices.workers.dev/library/research-memento-3): arXiv API record + arXiv HTML full text (read relevant sections). Could not verify: Private validation / semi-private score not reported ('not yet available to us'). No token counts. Per-game cost only as subscription-limit scale. Earlier harness versions leaked game identity (web search, second game client as simulator); those results excluded from reported scores. Development mostly on public game ls20.","url":"https://arxiv.org/abs/2605.05138","source":"arxiv","published":"2026-05-06","file":"corpus/research-memento-3/arxiv/2605.05138.pdf","sha256":"dbec3124b0e7549999ac58aa35c9132cd70ae671af0a07eb7d15c32f97582cd6","by":"owner","method":"research:memento-3","at":1791712838650,"changed":{"by":"owner","method":null,"at":1794680969746},"cites":[],"tags":["memento-3","agent-research"],"public":true,"staleSince":null,"wordsAt":1791713969439,"snippet":" … A [world]-[model] verifier replays recorded observations; a planner verifier checks [plans] reach LEVEL_COMPLETED in the [model]; a [plan] executor simulates a [plan] … "},{"id":"arxiv:2405.15383","kind":"paper","title":"Generating Code World Models with Large Language Models Guided by Monte Carlo Tree Search","text":"In this work we consider Code World Models, world models generated by a Large Language Model (LLM) in the form of Python code for model-based Reinforcement Learning (RL). Calling code instead of LLMs for planning has potential to be more precise, reliable, interpretable, and extremely efficient. However, writing appropriate Code World Models requires the ability to understand complex instructions, to generate exact code with non-trivial logic and to self-debug a long program with feedback from unit tests and environment trajectories. To address these challenges, we propose Generate, Improve and Fix with Monte Carlo Tree Search (GIF-MCTS), a new code generation strategy for LLMs. To test our approach in an offline RL setting, we introduce the Code World Models Benchmark (CWMB), a suite of program synthesis and planning tasks comprised of 18 diverse RL environments paired with corresponding textual descriptions and curated trajectories. GIF-MCTS surpasses all baselines on the CWMB and two other benchmarks, and we show that the Code World Models synthesized with it can be successfully used for planning, resulting in model-based RL agents with greatly improved sample efficiency and inference speed.\n\nMechanism: GIF-MCTS builds a Monte Carlo tree where nodes are partial programs (state = parent's code plus L=2 further lines; rollout = the remaining completion needed to run it) and edges are LLM actions. Selection uses UCT (with a custom exploration/exploitation heuristic, Appendix B); expansion calls the LLM with one of three actions; evaluation runs the rollout against unit tests (for a Code World Model, the fraction of transitions in 10 curated offline trajectories it predicts correctly) and the value is backpropagated. Three actions: (1) Generate new lines - LLM continues the node's code to a full program in a new branch; (2) Improve predictions - LLM is given the full parent program plus one wrongly predicted example (state, action, true next state, predicted next state) and asked for a chain-of-thought diagnosis and a rewrite; (3) Fix bugs - for programs that crash (reward 0) the LLM gets the error and fixes it; buggy nodes get a temporary value so UCT revisits them until fixed or attempts run out. Gate: program accuracy on the offline transitions (0 for non-executing code); search stops on matching all transitions or exhausting the LLM-call budget, best CWM returned.\n\nEvidence: CWMB (18 RL environments, discrete and continuous), Table 2, accuracy / normalized return R. Llama 3 70B (3 seeds, budget 50 LLM calls): GIF-MCTS discrete 0.84+/-0.03 acc, 0.76+/-0.03 R; continuous 0.35+/-0.03, 0.22+/-0.01. WorldCoder* (re-implementation): discrete 0.79+/-0.04, 0.60+/-0.04; continuous 0.32+/-0.03, 0.19+/-0.01. GPT-4 Turbo (1 seed, budget 10): GIF-MCTS discrete 0.91+/-0.08, 0.81+/-0.06; continuous 0.40+/-0.03, 0.26+/-0.01. WorldCoder* discrete 0.87+/-0.09, 0.79+/-0.06; continuous 0.24+/-0.06, 0.20+/-0.01. APPS competition split pass@20 (Llama 3 70B): GIF-MCTS 28.3+/-1.4, WorldCoder* 25.1+/-1.4, zero-shot CoT* 23.2+/-1.3. RTFM (simplified): GPT-4 Turbo budget 50 GIF-MCTS accuracy 1.00, R 1.00 vs WorldCoder* 0.64, R -0.06. Claim: planning with CWMs is 'from four to six orders of magnitude faster compared to directly querying an LLM as a world model'; only 10 trajectories of at most 100 steps used.\n\nCost: Budget is counted in LLM calls: 50 for Llama 3 70B, 10 for GPT-4 Turbo on CWMB (extra GPT-4 Turbo runs at 50 only on RTFM, 'we couldn't do this on the full CWMB due to budget concerns'). No dollar or wall-clock figures for synthesis found in the text.\n\nCode: https://github.com/nicoladainese96/code-world-models\n\nRead for the report on Memento 3 (https://cb4.endpointservices.workers.dev/library/research-memento-3): arXiv API record + arXiv HTML full text (read relevant sections). Could not verify: Dollar/token/wall-clock cost of synthesis not found. Note GPT-4 Turbo CWMB row is a single seed and the discrete-action gap to WorldCoder* is within the reported error bars (0.91+/-0.08 vs 0.87+/-0.09). Accuracy always exceeds normalized return (accuracy-return gap); Llama 3 failed to produce an executable CWM for Humanoid-v4 and HumanoidStandup-v4. WorldCoder baseline is the authors' own re-implementation of the first version of WorldCoder. CQL beats CWMs on continuous-action environments (Appendix K, not read in detail).","url":"https://arxiv.org/abs/2405.15383","source":"arxiv","published":"2024-05-24","file":"corpus/research-memento-3/arxiv/2405.15383.pdf","sha256":"64c9d0504344661f3477eebd0d92e1290da36f173805cef2d79a571152bf67f2","by":"owner","method":"research:memento-3","at":1791712838650,"changed":{"by":"owner","method":null,"at":1794680969746},"cites":[],"tags":["memento-3","agent-research"],"public":true,"staleSince":null,"wordsAt":1791713969439,"snippet":" … Calling code instead of LLMs for [planning] has potential to be more precise, reliable, interpretable, and extremely efficient. However, writing appropriate Code [World] [Models] … "},{"id":"arxiv:2610.26855","kind":"paper","title":"Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience","text":"World models enable embodied agents to anticipate action consequences, yet jointly modeling locomotion and manipulation across diverse embodiments remains challenging due to heterogeneous action spaces and limited multimodal supervision. We present Ropedia W0, the first egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. We curate Xp-Data, comprising 10k hours of egocentric human data and 3k hours of robot data with unified multimodal annotations and systematic quality control. A unified kinematic representation encodes locomotion through camera poses and manipulation through projected skeletons, with robot skeletons combining embodiment-specific forward kinematics with extensive manual keypoint annotation and correction. We also introduce Xp-Teleop, a long-horizon loco-manipulation benchmark based on high-quality dexterous humanoid teleoperation data. Evaluations demonstrate state-of-the-art performance across five cross-embodiment benchmarks (DreamDojo, Ψ_0, ω-HOME, Xp-Teleop, and AgiBot) and on held-out human egocentric video, from zero-shot inference to few-shot and full-data post-training. Scaling human pre-training data improves performance on both human and robot evaluations. After distillation, our model supports real-time, minute-long rollouts, enabling live VR interaction. Applications in policy evaluation, model-based planning, and world-action modeling further demonstrate the model's utility for embodied decisio","url":"https://arxiv.org/abs/2610.26855","source":"arxiv","published":null,"file":null,"sha256":null,"by":"worker:brain-x-snapshot","method":"snapshot","at":1792562443715,"changed":{"by":"owner","method":null,"at":1794680965863},"cites":[],"tags":["votes-hf-papers"],"public":true,"staleSince":null,"wordsAt":1792562443715,"snippet":" … Applications in policy evaluation, [model]-based [planning], and [world]-action [modeling] further demonstrate the [model]'s utility for embodied decisio"},{"id":"arxiv:2610.30120","kind":"paper","title":"ALPS: Visual State Encoders for World Models","text":"World models turn visual representations into predictions and actions. We argue that their visual foundation should combine physical-state information, native language alignment, and pixel reconstruction, so that a single representation can support prediction, instruction, and inspection. However, existing encoders typically specialize in only one or two of these capabilities; we therefore introduce ALPS, a pretrained visual state encoder that unifies all three in a single set of patch tokens. Using only still images and text, we pretrain a ViT-B/16 from scratch with language, reconstruction, and spatial correspondence objectives, and then refine it via cross-view self-distillation. In a unified evaluation across fifteen public encoders, ALPS achieves the lowest aggregate LIBERO state-probe error among ViT-B models (matching the 3.5x larger V-JEPA 2-L) while maintaining strong language alignment and accurate pixel reconstruction. Downstream, frozen ALPS tokens reach the highest mean policy success on LIBERO (85.5%) when paired with our native text tower, match or exceed leading encoders on world-model planning across PushT, Wall, and Granular, and allow predicted future tokens to be decoded directly into pixels for visual inspection. We hope this work helps establish visual pretraining as a shared foundation for understanding, prediction, and action in world models.","url":"https://arxiv.org/abs/2610.30120","source":"arxiv","published":null,"file":null,"sha256":null,"by":"worker:brain-x-snapshot","method":"snapshot","at":1792648868893,"changed":{"by":"owner","method":null,"at":1794680965863},"cites":[],"tags":["votes-hf-papers"],"public":true,"staleSince":null,"wordsAt":1792648868893,"snippet":" … our native text tower, match or exceed leading encoders on [world]-[model] [planning] across PushT, Wall, and Granular, and allow predicted future tokens to … "},{"id":"arxiv:2611.08449","kind":"paper","title":"VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding","text":"Models for physical AI tasks, such as robotics and autonomous driving, need not reason solely in language. They can use visual plans--points, bounding boxes, and trajectories--to represent spatial goals and anticipate motion. However, autoregressive vision-language models generate these plans one coordinate token at a time, making richer predictions increasingly costly. We introduce VFlash, a method for fast, accurate visual planning through Flash Primitive Decoding. VFlash exploits the known structure of visual primitives to predict their coordinates jointly, using a lightweight drafter that generates an entire primitive in one forward pass. The layout adapts to the primitive's type, dimensionality, and length, supporting compact boxes and long, variable-length trajectories. Crucially, spatial accuracy need not require reproducing an autoregressive model's exact token sequence. VFlash therefore trains on annotated primitives and directly accepts structurally valid predictions without target-token agreement. Across the evaluated 2D and 3D robot trajectory tasks, VFlash consistently reduces prediction errors relative to autoregressive decoding. Preliminary single-rollout measurements indicate approximately 25x faster generation for 3D point motion forecasting. A world action model conditioned on VFlash plans achieves 70.2% average success across four MolmoSpaces task suites, compared with 62.6% for the baseline.","url":"https://arxiv.org/abs/2611.08449","source":"arxiv","published":null,"file":null,"sha256":null,"by":"worker:brain-x-snapshot","method":"snapshot","at":1794031203859,"changed":{"by":"owner","method":null,"at":1794680960751},"cites":[],"tags":["votes-hf-papers"],"public":true,"staleSince":null,"wordsAt":1794031203859,"snippet":" … A [world] action [model] conditioned on VFlash [plans] achieves 70.2% average success across four MolmoSpaces task suites, compared with 62.6% for the … "},{"id":"arxiv:2607.28287","kind":"paper","title":"Tycho: Active Abstraction with Programmatic World Models for ARC-AGI-3","text":"ARC-AGI-3 turns abstraction into an interactive problem of skill acquisition. A player must infer an unfamiliar game's rules, hidden state, and goal while maintaining action efficiency because every move counts. We formalize these environments as parameterized rendered deterministic Moore machines and introduce Tycho, a coding-agent system that constructs and uses game-specific models during interaction. Tycho separates actionable observations from intermediate animation, level-completion, and game-over frames. From this structured history, an agent can model, test, plan with, repair, or bypass a free-form executable hypothesis.\n  In one matched public-set run per policy, we compare four orchestration policies on all 25 public games using Claude Opus 4.8 under matched inference budgets. Actor-requested delegation to a model builder obtains the highest observed mean Relative Human Action Efficiency (RHAE), 88.49. With this selected policy, GPT-5.6 Sol and Opus 5 both reach 100.00 RHAE and complete all 183 levels. Their game-balanced first-run human-replay midranks are 98.5 and 100.0. Opus 5 uses 61% fewer scored actions than the aggregate official human baselines.\n  Automatic repair after verification failures produces models that reproduce observed transitions much more accurately, yet reaches only 83.07 RHAE. Transition match indicates whether a simulator reproduces observed dynamics, not whether it has identified the objective or improves the next action. Strong play also requires deciding when to construct, repair, use, or bypass a model. We call this joint problem active abstraction: generating a testable model from costly interaction and deciding when acquiring or using it is worth its cost.\n\nMechanism: Coding-agent system (Python) with typed durable evidence (decision frames separated from transient, terminal, fatal, reset frames; wmlib library; per-level files) and a free-form executable world model the agent may build, verify, plan with, repair or bypass. Verification replays the latest attempt per level from its first frame and compares predicted to recorded successors cell by cell (accuracy on claimed pixels plus coverage); outcomes (level_complete/game_over) verified separately. Planning searches the program for a route to level_complete and writes a validated, hash-anchored artifact that the actor treats as conditional advice. Four orchestration policies compared: no world model, single (actor-authored), orchestrator (actor-requested delegation to a builder), trigger (automatic builder call on verification failure).\n\nEvidence: Claude Opus 4.8, 25 public games, one run per policy (RHAE): no world model 79.07, single 85.36, orchestrator 88.49, trigger 83.07; games completed 19/19/21/18, levels 157/162/166/162. Selected policy: GPT-5.6 Sol 100.00 RHAE (183 levels, 7,766 scored actions) and Opus 5 100.00 (6,641 actions).\n\nCost: Opus 4.8 list-price totals about USD 5.66k (no model), 7.27k (single), 5.78k (orchestrator), 8.03k (trigger); GPT-5.6 Sol run USD 4.47k total (mean 179, median 114 per game, max 649); Opus 5 USD 2.99k total (mean 119, median 97 per game, max 303), per the paper's estimates. A USD 1.5k per-game ceiling nominally applied.\n\nCode: https://github.com/NIMI-research/Tycho\n\nRead for the report on Memento 3 (https://cb4.endpointservices.workers.dev/library/research-memento-3): arXiv API record + arXiv HTML full text (read relevant sections). Could not verify: The paper itself says end-of-run measurements cannot determine how earlier model versions affected action; one run per policy, no variance. Causal claim rests on 4 policies that differ in several ways at once. Cost figures are the paper's list-price estimates, not billed.","url":"https://arxiv.org/abs/2607.28287","source":"arxiv","published":"2026-07-30","file":"corpus/research-memento-3/arxiv/2607.28287.pdf","sha256":"6e3b72ecee72ea3e2ecb125f4a40fcdbf67948164b0159e3f8778242a0428506","by":"owner","method":"research:memento-3","at":1791712838650,"changed":{"by":"owner","method":null,"at":1794680969746},"cites":[],"tags":["memento-3","agent-research"],"public":true,"staleSince":null,"wordsAt":1791713969439,"snippet":" … per-level files) and a free-form executable [world] [model] the agent may build, verify, [plan] with, repair or bypass. Verification replays the latest … "},{"id":"arxiv:2610.11794","kind":"paper","title":"Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks","text":"Learning to act in unfamiliar environments requires agents to infer how the world works and revise that understanding as new evidence arrives. Yet limited observations can support multiple world models that explain past interactions but predict different outcomes in unseen states. We introduce Memento 3, building on the Memento series to enable frozen LLM agents to continually learn explicit world models through external memory. The agent maintains a natural-language rulebook as persistent semantic memory, recording revisable hypotheses about environment dynamics while leaving unknown aspects underspecified. It compiles this rulebook into executable code for prediction and planning. Through a continual loop of observation, reflection, rule revision, compilation, and verification, the agent uses prediction errors to refine both the rulebook and its code. Updated code is accepted only when the LLM judges it faithful to the rulebook and cell-exact replay reproduces the observed transitions. We investigate this process as a model-based route to recursive self-improvement (RSI): the agent autonomously explores the environment, revises its world model, and uses verified updates to guide subsequent interaction and learning, while the underlying LLM remains fixed. A population extension maintains multiple world models in parallel, sharing interaction evidence and using their predictions to guide exploration. On ARC-AGI-3, the single-model agent clears every level of all 25 public gam","url":"https://arxiv.org/abs/2610.11794","source":"arxiv","published":null,"file":null,"sha256":null,"by":"worker:brain-x-snapshot","method":"snapshot","at":1791612048537,"changed":{"by":"owner","method":null,"at":1794680974470},"cites":[],"tags":["votes-hf-papers"],"public":true,"staleSince":null,"wordsAt":1791612048537,"snippet":" … We investigate this process as a [model]-based route to recursive self-improvement (RSI): the agent autonomously explores the environment, revises its [world] [model] … "},{"id":"arxiv:2610.20442","kind":"paper","title":"Towards Agentic Robots: Model, Data, Environment, and Harness","text":"Foundation models are driving a paradigm shift in robotic intelligence, moving beyond low-level action prediction to goal-directed orchestration of data collection, model training, task planning, and real-world execution. We characterize this emerging paradigm as an agentic robots, in which agents reason over human instructions, interact with robotic environments, rectify execution errors autonomously, and dynamically adjust operational strategies. Although this field develops fast, current studies are not well connected. Researchers have no consistent view of how to combine different agentic functions into a complete robot system. This paper provides a unified overview of existing research on agentic robots. We first trace the development from robot policies to agentic systems, covering vision-language-action models, world models, and world-action models. We then clarify the definition and scope of agentic robots and organize existing research into seventeen categories across four functional domains: Model, Data, Environment, and Harness. Across these domains, we examine how agents shape model learning and robot decisions, collect and curate robot data, construct worlds, tasks, and tests for learning and evaluation, and organize capabilities for robot use. Finally, we discuss open challenges and outline directions for future research. Looking forward, we hope this work contributes to a clearer understanding of agentic robots and supports new approaches to making them more ca","url":"https://arxiv.org/abs/2610.20442","source":"arxiv","published":null,"file":null,"sha256":null,"by":"worker:brain-x-snapshot","method":"snapshot","at":1792044058141,"changed":{"by":"owner","method":null,"at":1794680965863},"cites":[],"tags":["votes-hf-papers"],"public":true,"staleSince":null,"wordsAt":1792044058141,"snippet":" … prediction to goal-directed orchestration of data collection, [model] training, task [planning], and real-[world] execution. We characterize this emerging paradigm as an agentic … "},{"id":"arxiv:2402.12275","kind":"paper","title":"WorldCoder, a Model-Based LLM Agent: Building World Models by Writing Code and Interacting with the Environment","text":"We give a model-based agent that builds a Python program representing its knowledge of the world based on its interactions with the environment. The world model tries to explain its interactions, while also being optimistic about what reward it can achieve. We define this optimism as a logical constraint between a program and a planner. We study our agent on gridworlds, and on task planning, finding our approach is more sample-efficient compared to deep RL, more compute-efficient compared to ReAct-style agents, and that it can transfer its knowledge across environments by editing its code.\n\nMechanism: WorldCoder keeps a replay buffer D of (s,a,r,s',c,d) tuples plus each episode's initial state and context, and a Python program (transition function T-hat, goal-conditioned reward function R-hat) as the world model over symbolic object-set states. Loop: act by planning in the current code model (planner is separate; MCTS used for AlfWorld), store the experience, and if the program fails a constraint, prompt GPT-4 with the erroneous program plus transitions it cannot explain (and, if needed, a random subset of D) to fix its code; repeat. Gate: a program is only accepted when it satisfies phi_1 (explains every stored transition and reward) AND phi_2 (optimism: given the initial state and goal, there exists an action sequence under the model reaching a positive-reward terminal state). Which program to refine next is chosen by REx (a bandit balancing refining promising vs rarely-refined programs) with heuristic h = fraction of replay buffer consistent with phi_1 plus (indicator phi_1) x fraction of (s0,c) pairs satisfying phi_2. Transfer across environments = refining an old program rather than writing a new one.\n\nEvidence: Sokoban: starting from only the goal 'win the game', the agent builds a world model over the first 50 actions; ReAct succeeds on only 15% +/- 8% of basic levels; deep RL (model-based and model-free) requires millions of experiences / >1 million steps for 2-box Sokoban (Fig 3D); cannot solve 5+ box levels (planner limitation). Optimism (phi_2) has little effect on Sokoban (dense reward). Minigrid: PPO needs 10^4-10^5 steps to learn valid policies in most environments and cannot solve UnlockPickup even in 3x10^8 steps; phi_2 is necessary for harder environments without transfer and for zero-shot new goals; 'prompt4reward' ablation struggles to follow new instructions. AlfWorld: synthesizes a 250+ line world model; optimism objective necessary for nonzero performance; typically solves the task in the first episode after about 20 exploratory steps. Abstract: more sample-efficient than deep RL, more compute-efficient than ReAct-style agents, transfers knowledge by editing code.\n\nCost: 'after front-loading 400k LLM tokens (about $15), it can issue as many actions as needed without subsequent LLM queries' (Sokoban). Model: GPT-4 for refinement. Asymptotic LLM calls per new task O(1) vs O(T) for ReAct-style agents.\n\nRead for the report on Memento 3 (https://cb4.endpointservices.workers.dev/library/research-memento-3): arXiv API record + arXiv HTML full text (read relevant sections). Could not verify: No code repository URL found in the text (NeurIPS checklist says 'The paper has not released new assets'). Exact numeric Sokoban/Minigrid success rates are only in figures, not in text, other than the ReAct 15% +/- 8% figure. No wall-clock time stated. Limitations stated: deterministic dynamics only, symbolic discrete states required.","url":"https://arxiv.org/abs/2402.12275","source":"arxiv","published":"2024-02-19","file":"corpus/research-memento-3/arxiv/2402.12275.pdf","sha256":"55c47b1df146967ecab1933518783059885b693460d3ebbc7554aaec0eeb101f","by":"owner","method":"research:memento-3","at":1791712838650,"changed":{"by":"owner","method":null,"at":1794680969746},"cites":[],"tags":["memento-3","agent-research"],"public":true,"staleSince":null,"wordsAt":1791713969439,"snippet":" … R-hat) as the [world] [model] over symbolic object-set states. Loop: act by [planning] in the current code [model] (planner is separate; MCTS … "},{"id":"arxiv:2611.12324","kind":"paper","title":"How Many Exposures Does a World Model Need? Learning Unseen Game Mechanics at Test Time","text":"World models enable agents to plan and learn from imagined experience, but in an unfamiliar environment they must learn its mechanics from limited interaction. A mechanic is a rule for how an action or an interaction changes part of the environment. Each occurrence of a mechanic that the model observes is one exposure. We ask how many exposures a world model needs to predict the effects of a new environment's mechanics. We address this question on ARC-AGI-3, a benchmark of interactive games defined by their programs. We introduce MechEval, which labels 69 recurring mechanics across 20 games from these programs. It scores each prediction on the cells the mechanic changes at each exposure. As a test bed, we build WARC (World model for ARC-AGI-3), a 35M-parameter vision transformer that is pretrained on other games and updates all of its parameters after each transition. On held-out games, WARC learns to predict four in five mechanics with at least 80% accuracy on the cells they change and keeps that accuracy to the end of the recordings. Half of them reach that accuracy within about 80 exposures, in a single play of the game, and without updates only one in seven does at all. In closed-loop play on the public games, WARC-Agent, which pairs WARC with a fixed graph explorer, triples the explorer's official efficiency score from 0.15 to 0.46, mostly through test-time training. On the hidden games of the Kaggle leaderboard, our entry scores 0.20.","url":"https://arxiv.org/abs/2611.12324","source":"arxiv","published":null,"file":null,"sha256":null,"by":"worker:brain-x-snapshot","method":"snapshot","at":1794376812031,"changed":{"by":"owner","method":null,"at":1794680960751},"cites":[],"tags":["votes-hf-papers"],"public":true,"staleSince":null,"wordsAt":1794376812031,"snippet":"[World] [models] enable agents to [plan] and learn from imagined experience, but in an unfamiliar environment they must learn its mechanics from limited interaction … "},{"id":"arxiv:2608.14490","kind":"paper","title":"Twin: Playing an Unknown Game with a Test-Time Digital Twin","text":"We present a Test-time World-model Inference (Twin) system, in which a frontier coding agent writes an executable world model for completing continual learning tasks, such as ARC-AGI-3 games. Traditional approaches hand-engineer such models, one custom design per task. Each game hides its rules and goal, and our system constructs them from simulation and interaction alone. Its inductive prior over grid games is strong enough to recover the true transitions of the game and the goal on nearly all levels. Replay validation happens in a twin world model. The harness enforces that an action is not made until the program reproduces every previous observed game transition. Each mismatch between a world model prediction and the actual action result becomes a counterexample that is used to repair the world model. Twin clears 179 out of 183 levels (97.8%), and does so more efficiently than humans in 158 out of 179 levels (88.3%). The system infers the goal before any reward on 156 of the levels it clears (87.2%), and in the remaining levels automatically discovers the goal by search. The benchmark scores completion and action efficiency, between 0 and 100, against humans playing each game for the first time. Played directly, the base model scores only 7.8%; an off-the-shelf harness increases it to 61.1%, whereas our twin world model increases the same base model to 93.3%, clearing 23 out of 25 games. Building a usable world model is simpler than anticipated, whereas the harder problem is inferring the right goal.\n\nMechanism: Per game, a frontier coding agent (OpenAI Codex, GPT-5.6 Sol, connected to the game only through files) starts from an identity-twin Python world model (T,R) and an empty log D of all observed transitions. Loop per move: Perceive the real grid; while Validate(T,D) returns gaps (twin fails to replay any logged transition), Explore ranks broken rules and Repair rewrites T (dynamics wall); then Plan by BFS inside the twin; if no plan, Explore ranks reachable candidates and ProposeGoal sets a tentative R (goal wall). Acceptance gates enforced by the harness: no scored action is issued until the twin reproduces every logged transition; a tentative goal must evaluate false on every logged frame; a candidate reached without a level boundary is never proposed again; actions are submitted under a halt-on-mismatch check, with each mismatch becoming a counterexample for repair.\n\nEvidence: ARC-AGI-3 public set (25 games, 183 levels). Twin scores 93.3, clears 23/25 games and 179/183 levels (97.8%), more efficiently than humans on 158 of 179 cleared levels; goal inferred before any reward on 156 of cleared levels (87.2%). Same base model played directly scores 7.8%; off-the-shelf Codex harness (no-twin ablation) 61.1 (13 games, 148 levels). Table 1: OPINE-World 78.4 (20 games, 160 levels), Prime Agent 78.3 with GPT-5.6 Sol (164 levels), EWM 63.8 with GPT-5.5 (14 games, 146 levels), Codex GPT-5.6 Sol 61.1. Twins predict 79% of first-seen state-action pairs exactly. 18 games reach 100.0.\n\nCost: Twin used 2.60 billion processed tokens (47.8M non-cached) and 91.4 hours wall-clock inference across the 25 runs; roughly 224k tokens per scored action; per-game range 5.1M (sb26) to 625.3M (ka59) tokens; ka59 alone 24% of tokens. No-twin ablation: 1.06B tokens, 48k tok/action, 22,224 actions vs Twin 11,614 actions. EWM: 20.0B tokens, 715k tok/action. No dollar figure.\n\nRead for the report on Memento 3 (https://cb4.endpointservices.workers.dev/library/research-memento-3): arXiv API record + arXiv HTML full text (read relevant sections). Could not verify: No dollar cost, only tokens and hours; per-task cost expressed as tokens (224k/action; per-game 5.1M-625M). Not checked whether scores are single runs per game. Limitations stated: deterministic dynamics only, replay certifies only logged transitions, fixed search budgets, no long-temporal-context mechanics. Negative: Twin uses 2.4x total tokens and 4.7x tokens per action vs the no-twin ablation; tn36 (1.5x human) and ka59 (1.6x) need more actions than humans; OPINE-World ahead on g50t, sp80, sc25; two games not cleared. | code note: https://github.com/Alexyskoutnev/TWIN-ARC-AGI-3 (site: https://arc-agi-3-twin.vercel.app/)","url":"https://arxiv.org/abs/2608.14490","source":"arxiv","published":"2026-08-14","file":"corpus/research-memento-3/arxiv/2608.14490.pdf","sha256":"ea425e1ffde0403aba56b55f1d74866b8c88c2028eaaededd36c9361de440e0c","by":"owner","method":"research:memento-3","at":1791712838650,"changed":{"by":"owner","method":null,"at":1794680969746},"cites":[],"tags":["memento-3","agent-research"],"public":true,"staleSince":null,"wordsAt":1791713969439,"snippet":"We present a Test-time [World]-[model] Inference (Twin) system, in which a frontier coding agent writes an executable [world] [model] for completing continual … "},{"id":"arxiv:2410.07484","kind":"paper","title":"WALL-E: World Alignment by Rule Learning Improves World Model-based LLM Agents","text":"Can large language models (LLMs) directly serve as powerful world models for model-based agents? While the gaps between the prior knowledge of LLMs and the specified environment's dynamics do exist, our study reveals that the gaps can be bridged by aligning an LLM with its deployed environment and such \"world alignment\" can be efficiently achieved by rule learning on LLMs. Given the rich prior knowledge of LLMs, only a few additional rules suffice to align LLM predictions with the specified environment dynamics. To this end, we propose a neurosymbolic approach to learn these rules gradient-free through LLMs, by inducing, updating, and pruning rules based on comparisons of agent-explored trajectories and world model predictions. The resulting world model is composed of the LLM and the learned rules. Our embodied LLM agent \"WALL-E\" is built upon model-predictive control (MPC). By optimizing look-ahead actions based on the precise world model, MPC significantly improves exploration and learning efficiency. Compared to existing LLM agents, WALL-E's reasoning only requires a few principal rules rather than verbose buffered trajectories being included in the LLM input. On open-world challenges in Minecraft and ALFWorld, WALL-E achieves higher success rates than existing methods, with lower costs on replanning time and the number of tokens used for reasoning. In Minecraft, WALL-E exceeds baselines by 15-30% in success rate while costing 8-20 fewer replanning rounds and only 60-80% of tokens. In ALFWorld, its success rate surges to a new record high of 95% only after 6 iterations.\n\nMechanism: World model = frozen LLM + a learned rule set (natural-language rules R^NL plus code rules R^code). Each iteration the agent (MPC, look-ahead planning with the world model) collects a real trajectory; predicted transitions are compared to real ones on the binary action_result and split into correct / incorrect sets. An LLM induces new NL rules from real trajectories (distinct from previous rules), then refines existing rules (modifies/discards conflicting ones), then translates NL rules to code. Gate: code rules are executed against the labeled predicted transitions and any rule that fails to predict a transition correctly is discarded; the survivors are pruned by a greedy maximum-coverage selection over the transitions the bare LLM got wrong (rules covering only correct transitions or fully redundant are dropped, and dropped code rules are also removed from the NL set). At inference NL rules go in the prompt; code rules override the LLM prediction as a verification layer (ApplyRules). Gradient-free.\n\nEvidence: Minecraft (Table 1, success rate %, avg): WALL-E 69 (15.77 replanning rounds) vs GITM 54 (25.49), DEPS 37 (35.36), WALL-E w/o WM 61 (23.13), Optimus-1* 47, Jarvis-1* 42, Human 59. Per level WALL-E: Wooden 98, Stone 91, Iron 63, Golden 69, Diamond 46, Redstone 48. ALFWorld (Table 3, 134 test tasks, avg): WALL-E 95; RAFA 95; AdaPlanner 91; Reflexion 86; ReAct 74; EMMA 82; Human 91. Abstract: 'In Minecraft, WALL-E exceeds baselines by 15-30% in success rate while costing 8-20 fewer replanning rounds and only 60-80% of tokens. In ALFWorld, its success rate surges to a new record high of 95% only after 6 iterations.' Ablation (Table 4): LLM agent + LLM world model 38(33.53); rules in world model 69(15.77); rules in agent only 61(23.13); rules in both 67(16.59). World model without rules barely helps. Rule set stops changing after 4 iterations.\n\nCost: Table 2 Minecraft avg tokens (USD): WALL-E 60348.71 (0.41), GITM 74638.54 (0.51), DEPS 93560.95 (0.65), WALL-E w/o WM 72390.16 (0.52). GPT-4o backend for Minecraft; GPT-3.5-Instruct backbone for ALFWorld (one-step MPC). Minecraft 30 training / 24 testing tasks; ALFWorld 134 test tasks; 3 training tasks per iteration over 10 iterations in the rule-learning study. Per-task 10-minute limit.\n\nCode: https://github.com/elated-sawyer/WALL-E\n\nRead for the report on Memento 3 (https://cb4.endpointservices.workers.dev/library/research-memento-3): arXiv API record + arXiv HTML full text (read relevant sections). Could not verify: The paper has no 'Mars' benchmark; its benchmarks are Minecraft (MineDojo-style tasks) and ALFWorld. Did not find a stated wall-clock total for rule learning or a per-iteration LLM cost. In ALFWorld, WALL-E's 95 average ties RAFA's 95 in Table 3 despite the 'new record high' phrasing (per-category: WALL-E 100/97/100/86/85/100 vs RAFA row partly read; RAFA per-category not fully read). ALFWorld token cost not found.","url":"https://arxiv.org/abs/2410.07484","source":"arxiv","published":"2024-10-09","file":"corpus/research-memento-3/arxiv/2410.07484.pdf","sha256":"e49c6e1c07f30a9ca0bc84d2a3489adf3314e4851539a8ab40df4c453b2fcb21","by":"owner","method":"research:memento-3","at":1791712838650,"changed":{"by":"owner","method":null,"at":1794680969746},"cites":[],"tags":["memento-3","agent-research"],"public":true,"staleSince":null,"wordsAt":1791713969439,"snippet":" … Each iteration the agent (MPC, look-ahead [planning] with the [world] [model]) collects a real trajectory; predicted transitions are compared to real ones on … "},{"id":"arxiv:2610.31494","kind":"paper","title":"AgentVQA: A Unified Benchmark for Agentic Visual Understanding","text":"Vision-language models (VLMs) can perform a broad range of tasks across diverse settings, yet their performance in agentic contexts remains poorly understood. Although online evaluations in simulators and real-world environments are essential for measuring interactive agent performance, they are costly and challenging to scale. Complementary offline VQA benchmarks can enable faster and more reliable comparisons across agentic domains, yet existing benchmarks remain limited and fragmented. To address this gap, we introduce AgentVQA, an offline benchmark for evaluating VLMs on realistic agentic scenarios across five agent domains: Web Agents, Robotics, Egocentric Videos, Games, and Spatial Planning. We build AgentVQA by manually selecting 1000 challenging, real-world agent examples from 23 source datasets and adapting them for reliable offline evaluation. The resulting benchmark preserves realistic scenarios from interactive environments across diverse domains, and its model rankings track those from corresponding online evaluations. Across the 22 evaluated models, the best achieves only 58.5% accuracy. Our cross-domain analysis shows that errors are driven primarily by failures to ground and maintain accurate representations of the current and evolving visual state, including localizing relevant UI elements, interpreting motion, tracking state changes over time, and inferring 3D spatial relationships, rather than by task misinterpretation or downstream logical reasoning.","url":"https://arxiv.org/abs/2610.31494","source":"arxiv","published":null,"file":null,"sha256":null,"by":"worker:brain-x-snapshot","method":"snapshot","at":1792821678107,"changed":{"by":"owner","method":null,"at":1794680960751},"cites":[],"tags":["votes-hf-papers"],"public":true,"staleSince":null,"wordsAt":1792821678107,"snippet":" … Web Agents, Robotics, Egocentric Videos, Games, and Spatial [Planning]. We build AgentVQA by manually selecting 1000 challenging, real-[world] agent examples from 23 source … "},{"id":"arxiv:2611.14309","kind":"paper","title":"MMDM: Measuring Massive Multitask Decision Making","text":"System-1 Decision Models directly make structured probabilistic decisions over query-specified options given a state in a single forward pass, without autoregressively generating reasoning trajectories or textual responses. Recent evaluations cover diverse tasks but often rely on existing benchmarks, where prior exposure to test material can obscure generalization. Moreover, real-world decision making often involves long contexts, large option spaces, and complex logical dependencies, challenges that existing benchmarks do not adequately capture together. Thus, we introduce MMDM (Massive Multitask Decision Making), a benchmark of 5,448 high-quality questions across 16 task categories, combining newly constructed tasks with source-grounded recompositions and rewrites for single-select, multi-select, and soft-label decision making in fresh contexts. It covers realistic scenarios such as agent state tracking, tool and skill selection, preference judgments, and rubric scoring, with option spaces ranging from 2 to 1,024 and inputs extending to 64,817 tokens. With MMDM, we conduct a comprehensive quantitative evaluation of 28 decision models, showing that many perform well on common knowledge yet struggle with interdependent options, constrained decisions, and multi-select tasks. For example, Perplexity Decider-V1.2 reaches 78.75% Hard Single-select Accuracy on world knowledge, but only 53.95% on option logic and 47.95% on constraint planning, while the best Hard Multi-select Accur","url":"https://arxiv.org/abs/2611.14309","source":"arxiv","published":null,"file":null,"sha256":null,"by":"worker:brain-x-snapshot","method":"snapshot","at":1794376812031,"changed":{"by":"owner","method":null,"at":1794680960751},"cites":[],"tags":["votes-hf-papers"],"public":true,"staleSince":null,"wordsAt":1794376812031,"snippet":" … Single-select Accuracy on [world] knowledge, but only 53.95% on option logic and 47.95% on constraint [planning], while the best Hard Multi … "},{"id":"arxiv:2610.21453","kind":"paper","title":"CarryOn: A Dual-Loop Harness for Personalized and Reliable Robot Assistance","text":"While generalist vision-language-action (VLA) models have enabled diverse open-world manipulation, real-world robotic assistance requires two critical capabilities they currently lack: personalization and reliable execution. First, executing an instruction like \"make me a sandwich\" requires grounding subtask decomposition in persistent user preferences rather than relying on a single generic solution. Second, reliable deployment demands explainable safety assurance that proactively prevents hazards and gracefully recovers from occasional physical failures. To address these gaps, we present CarryOn, a dual-loop harness designed to augment a frozen VLA policy. The planning loop maintains persistent user-level and scene-level context that is continually grounded against new observations and revised through interaction, enabling personalized task selection and execution grounding across conversations and changing scenes. The execution loop actively supervises policy rollouts, shielding actions from predicted hazardous contact. When execution fails, the system invokes reusable control-based recovery skills that restore task progress and are refined through reflection on earlier failures. On a real robot and in simulation, CarryOn dispatches preference-consistent subtasks across users and scenes, reduces the violation rate of an already safety-finetuned VLA from 52.6% to 6.1%, and increases recovery success from 13% to 80% when refined skills are reused.","url":"https://arxiv.org/abs/2610.21453","source":"arxiv","published":null,"file":null,"sha256":null,"by":"worker:brain-x-snapshot","method":"snapshot","at":1792216816089,"changed":{"by":"owner","method":null,"at":1794680965863},"cites":[],"tags":["votes-hf-papers"],"public":true,"staleSince":null,"wordsAt":1792216816089,"snippet":"While generalist vision-language-action (VLA) [models] have enabled diverse open-[world] manipulation, real-[world] robotic assistance requires two critical capabilities they currently lack … "},{"id":"arxiv:2510.04542","kind":"paper","title":"Code World Models for General Game Playing","text":"Large Language Models (LLMs) reasoning abilities are increasingly being applied to classical board and card games, but the dominant approach -- involving prompting for direct move generation -- has significant drawbacks. It relies on the model's implicit fragile pattern-matching capabilities, leading to frequent illegal moves and strategically shallow play. Here we introduce an alternative approach: We use the LLM to translate natural language rules and game trajectories into a formal, executable world model represented as Python code. This generated model -- comprising functions for state transition, legal move enumeration, and termination checks -- serves as a verifiable simulation engine for high-performance planning algorithms like Monte Carlo tree search (MCTS). In addition, we prompt the LLM to generate heuristic value functions (to make MCTS more efficient), and inference functions (to estimate hidden states in imperfect information games). Our method offers three distinct advantages compared to directly using the LLM as a policy: (1) Verifiability: The generated CWM serves as a formal specification of the game's rules, allowing planners to algorithmically enumerate valid actions and avoid illegal moves, contingent on the correctness of the synthesized model; (2) Strategic Depth: We combine LLM semantic understanding with the deep search power of classical planners; and (3) Generalization: We direct the LLM to focus on the meta-task of data-to-code translation, enabling it to adapt to new games more easily. We evaluate our agent on 10 different games, of which 4 are novel and created for this paper. 5 of the games are fully observed (perfect information), and 5 are partially observed (imperfect information). We find that our method outperforms or matches Gemini 2.5 Pro in 9 out of the 10 considered games.\n\nMechanism: Before play, Gemini 2.5 Pro is given the game's natural-language rules plus 5 offline trajectories and writes a Python code world model (CWM) in OpenSpiel API format (state transition, legal moves, termination); for imperfect-information games it also writes an inference function (samples hidden state/history) and optionally a heuristic value function. Refinement: unit tests are auto-generated from every transition in the offline trajectories (states, observations, rewards, action legality, no execution errors); failing stack traces are fed back to the LLM. Two refinement modes: sequential conversation, or tree search a la REx with Thompson sampling over multiple CWMs (chosen for the rest of the paper). Refinement runs until transition accuracy is 1.0 or the LLM-call budget is exhausted. Value functions are not refined. For closed-deck (no hidden info) a 'CWM autoencoder' uses the inference function as encoder and the CWM as decoder. The CWM then drives (IS)MCTS (1,000 simulations).\n\nEvidence: 10 games: perfect-info Tic-tac-toe, Connect four, Backgammon, Generalized tic-tac-toe (OOD), Generalized chess (OOD); imperfect-info Leduc poker, Bargaining, Gin rummy, Quadranto (OOD), Hand of war (OOD). Abstract: 'our method outperforms or matches Gemini 2.5 Pro in 9 out of the 10 considered games.' Validation: held-out test set of 10,000 transitions from 100 games with random/MCTS players on ground-truth code, plus online transitions during play. CWM-MCTS ties GT-MCTS (ground-truth-code MCTS upper bound) and beats Gemini 2.5 Pro in all perfect-info games; CWM-ISMCTS beats or matches Gemini except Hand of war. Table 1 (open deck, tree search, hidden-history inference): Gin rummy transition accuracy train 0.7816 / test 0.7455 / online 0.9044, inference accuracy train 0.5857 / test 0.5376, 500.0 LLM calls; Hand of war inference test 0.9357; most others 1.0 train. Matches averaged over 100 games, CWM synthesis repeated 5 times with bad samples rejected.\n\nCost: Synthesis model Gemini 2.5 Pro. Cost reported as number of LLM calls per game (Table 1: e.g. Leduc poker 4.4, Quadranto 6.0, Bargaining 23.0, Hand of war 338.2, Gin rummy 500.0 = budget). Tree search uses fewer calls than conversation. No dollars or tokens found.\n\nRead for the report on Memento 3 (https://cb4.endpointservices.workers.dev/library/research-memento-3): arXiv API record + arXiv HTML full text (read relevant sections). Could not verify: No code repository URL found. Text in 5.1.2 says Gin rummy training accuracy 84% / test 79%, but Table 1 shows 0.7816 / 0.7455 (different tables or settings, unreconciled). The one game where CWM-ISMCTS does not match Gemini is Hand of war, despite CWM accuracy ~98%. Gin rummy: inference accuracy ~52% and Gemini forfeits a lot, so that win is weak evidence. No token or dollar cost. Games are fixed-rule board/card games with 5 offline trajectories of ground truth, not interactive unknown-goal environments.","url":"https://arxiv.org/abs/2510.04542","source":"arxiv","published":"2025-10-06","file":"corpus/research-memento-3/arxiv/2510.04542.pdf","sha256":"03ab77dca2b96d3cfa38edb605bb78721e16cbc5159391fdf07ff8b6bf696982","by":"owner","method":"research:memento-3","at":1791712838650,"changed":{"by":"owner","method":null,"at":1794680969746},"cites":[],"tags":["memento-3","agent-research"],"public":true,"staleSince":null,"wordsAt":1791713969439,"snippet":" … We use the LLM to translate natural language rules and game trajectories into a formal, executable [world] [model] represented as Python code. This generated … "},{"id":"arxiv:2610.36068","kind":"paper","title":"Rollback: Experience Backtracking in World Models for Policy Improvement","text":"Action-conditioned world models offer a scalable alternative to costly real-world interaction for improving robot policies. However, existing world-model-based methods largely treat world models as forward-only surrogates of the physical environment, leaving untapped a capability no physical environment affords: rewinding to past decision points to explore counterfactual futures. We introduce Rollback, a policy improvement paradigm that turns world models from passive simulators into counterfactual engines: it rolls back to critical decision points, branches into alternative actions, and contrasts their outcomes to attribute success to individual decisions. This concentrates imagination where it matters most and distills counterfactual outcomes into fine-grained, action-level supervision without a separately learned value model. We build a scalable training system that unifies distributed world-model search with asynchronous real-robot execution, forming a self-reinforcing cycle: real-world experience continually sharpens the world model, and a sharper world model in turn drives further policy improvement. We term the world-model side of this cycle test-time physics grounding, a capability that hand-built physics simulators lack: while the world model serves as the training environment, real transitions from the policy it trains correct its dynamics through a lightweight LoRA pipeline. Experiments on three challenging real-world tasks show that Rollback improves success rates","url":"https://arxiv.org/abs/2610.36068","source":"arxiv","published":null,"file":null,"sha256":null,"by":"worker:brain-x-snapshot","method":"snapshot","at":1793167220610,"changed":{"by":"owner","method":null,"at":1794680960751},"cites":[],"tags":["votes-hf-papers"],"public":true,"staleSince":null,"wordsAt":1793167220610,"snippet":"Action-conditioned [world] [models] offer a scalable alternative to costly real-[world] interaction for improving robot policies. However, existing [world]-[model]-based methods largely … "},{"id":"arxiv:2610.12299","kind":"paper","title":"Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction","text":"Egocentric world models predict first-person observations conditioned on an agent's actions, but most focus on a single agent. Real embodied settings often involve multiple agents that act and interact within a shared environment. Existing multi-agent world models rely on coarse actions like locomotion, camera control, or discrete commands, leaving fine-grained embodied interactions underexplored. We formulate multi-agent egocentric world modeling as synchronized ego-stream generation for multiple agents interacting through fine-grained actions in a shared world. This requires cross-view action consistency, shared-environment consistency, and consistent propagation of interaction-induced state updates. We propose Multi-agent Egocentric World Model (ME-World), which jointly denoises multiple ego streams in a shared token sequence, conditions each stream on all agents' target-view poses, and grounds generation with shared environment memory. We train and evaluate on real and synthetic multi-agent data and introduce shared-world consistency metrics for environment, update, and identity consistency. Experiments show ME-World improves shared-world consistency, action control, identity preservation, and video quality over existing methods.","url":"https://arxiv.org/abs/2610.12299","source":"arxiv","published":null,"file":null,"sha256":null,"by":"worker:brain-x-snapshot","method":"snapshot","at":1791528035366,"changed":{"by":"owner","method":null,"at":1794680974470},"cites":[],"tags":["votes-hf-papers"],"public":true,"staleSince":null,"wordsAt":1791528035366,"snippet":"Egocentric [world] [models] predict first-person observations conditioned on an agent's actions, but most focus on a single agent. Real embodied settings often … "}]}