robotics software teams

Abstract: We establish finite-time convergence guarantee for synchronous relative value iteration (RVI) $Q$-learning for finite weakly communicating Markov decision processes (MDPs). For stochastic Krasnosel'skiĭ--Mann iterations driven by a non-expansive, linearly regular map on a finite-dimensional normed space, we establish last-iterate mean-squared residual bounds under martingale-difference noise whose conditional second moment grows at most linearly with the squared residual. Suitable polynomial stepsizes of order $n^{-a}$ yield an $O(n^{-a})$ rate for $0<a\leq1$. We verify linear regularity and the noise condition for the centered Bellman operator for finite weakly communicating MDPs. Applying the general result gives an $O(n^{-1})$ bound on the mean-squared Bellman span residual of RVI $Q$-learning iterates when a suitable stepsize sequence is used. The positive gap between optimal and suboptimal stationary deterministic policies then yields an $O(n^{-1})$ bound on the expected suboptimality of the greedy policy. Consequently, $O(ε^{-1})$ synchronous iterations suffice for an $ε$-optimal policy in expectation. These guarantees are instance-dependent, with constants that depend on the hitting-time parameters and the suboptimality gap of the second-best policy.

Abstract: We introduce iMED, the first synchronized multi-endoscope dataset for robot-assisted minimally invasive surgery (RAMIS), designed to address a fundamental limitation in surgical vision benchmarking: the absence of independent held-out viewpoints. Existing surgical datasets capture single-trajectory sequences from a single endoscope, making it impossible to distinguish geometric generalization from photometric interpolation along a narrow forward-facing path. iMED provides 340 sequences ~170K synchronized timepoints with 4 views per timepoint) recorded simultaneously from two independent stereo endoscopes across ex vivo, postmortem, and live surgical settings, spanning 14 specimens with diverse anatomical regions and motion regimes. The dataset includes calibrated camera intrinsics, ArUco-based frame-wise pose estimates with uncertainty quantification, instrument segmentation masks, and rich clinical metadata. Using a train-on-one-endoscope, test-on-another evaluation protocol, we benchmark 23 state-of-the-art methods across rigid and deformable novel view synthesis, pose estimation, feature matching, and monocular depth estimation. Our experiments show that methods relying primarily on photometric supervision degrade substantially under this held-out-endoscope setting, while methods with explicit geometric regularization are more robust across views in our benchmark. Notably, large-scale foundation models transfer surprisingly well to surgical imagery, often outperforming domain-specific models, suggesting pitfalls in current fine-tuning protocols. iMED establishes a new evaluation protocol for geometric generalization in surgical vision.

Abstract: Recent multimodal agents have shown promising progress in embodied planning and web interaction. However, existing evaluations often rely on simplified tasks with limited combinatorial structure, leaving it unclear whether current agents can robustly integrate perception and reasoning for multi-goal planning under temporal, physical, and multimodal constraints. We introduce DualWorldBench, a unified benchmark for evaluating agent planning under controlled symbolic--temporal and grounded--topological complexity. Built around a shared pickup--delivery interface, DualWorldBench consists of three complementary components: TextWorld, which isolates symbolic planning under varying task compositions and release dynamics; SimWorld, which evaluates grounded path planning under controlled topological variation with text, image, or multimodal inputs; and DualWorld-VQA, a diagnostic suite spanning perception, scene understanding, and planning. Experiments reveal a common gap between local understanding and globally consistent planning. In TextWorld, agents struggle to derive optimal task sequences under interacting temporal constraints; in SimWorld, grounded observations are not reliably converted into feasible action sequences. DualWorld-VQA further confirms that perception and topology understanding do not necessarily translate into constrained planning. Together, these results identify the central challenge as composing symbolic or grounded evidence into globally feasible and efficient multi-goal plans. These findings establish DualWorldBench as a challenging testbed for agents with tighter perception--reasoning--planning integration. All datasets and code are publicly available at our project website https://dualworld.netlify.app/.

Abstract: End-to-end planners decode the ego trajectory either by coordinate regression, which is fast but blends plausible maneuvers into one trajectory that may follow none of them, or by diffusion and flow matching, which model the trajectory distribution but need several inference steps. Drifting models move this iteration into training: a one-step generator learns to follow an attraction--repulsion field between generated and ground-truth samples. Drifting was developed for class-conditional image generation, where every class supplies many positive samples. In planning, every scene is its own condition. Pooling scenes would attract the generated trajectories of one scene to the ground truth of unrelated scenes, whereas computing the field per scene leaves a single positive, the expert trajectory. Mutual repulsion then holds the generated trajectories on a shell around it, so any single generated trajectory deviates from the ground truth. We present DriftingPlanner, a trajectory planner trained with the drifting loss alone, without any coordinate regression term. It computes the field per scene and removes the shell at inference: it decodes every noise vector together with its negation in one batched pass, and averaging each antithetic pair cancels the linear part of the displacement from the shell center. On NAVSIM navtest, DriftingPlanner reaches a state-of-the-art 93.75 PDMS, and with single-step inference its drifting decoder comes within 0.13 PDMS of the best diffusion or flow matching setting, which needs ten inference steps.

Abstract: Mobile GUI agents are typically fine-tuned 7B-scale vision-language models (VLMs) that generate each action as text. However, these VLMs decode every action token by token, and even agents that score candidate actions keep a large language model or VLM in the loop, making them slow and memory-hungry. To address these issues, we propose GUI-JEPA, to our knowledge the first decoder-free GUI agent policy built on a Joint-Embedding Predictive Architecture (JEPA), which selects actions in the latent space of a JEPA encoder pretrained on unlabelled GUI screenshots. Its action- and region-conditioned latent predictor turns the screen latent into one latent per candidate action, letting lightweight heads score all candidates against the instruction in a single batched pass, without decoding a token. We further show that the scorer relies on this latent pathway and that the predictor is trained for the decision itself: a randomly initialised predictor does just as well as one pretrained on screen dynamics. On AndroidControl, GUI-JEPA is 2.8-4.0x faster on the same GPU than a Qwen2.5-VL-7B agent we fine-tune with the same inputs, while reaching 86-92% of its in-distribution accuracy with 6% of its parameters. Code is available at https://github.com/bcwang-sjtu/GUI-JEPA

Abstract: Can LLM agents carry a hard real-world problem from raw data to a system that holds its own against people? We study this in Hearthstone Battlegrounds, an eight-player auto-battler with stochastic shops and battles, hidden information and nearly a thousand interacting cards. We present AutoTavern, a self-improving research system: specialised LLM research agents share a typed research memory, every claim passes verification gates, and live games close the loop, so the agents improve the simulator, opponents and policy and, from the same feedback, their own research process; a human sets goals and rules on ambiguous mechanics. In 20 days AutoTavern built TavernGym, the first faithful, full-content reproduction of the live game, and delivered 23 versions of TavernAgent, trained with PPO against opponents calibrated to real games. TavernAgent plays complete games on the live ladder by itself: its rating rose from 6,000 to about 15,000 and, on the public leaderboard's measure, its last 20 records average a placement of 1.60 with 80% first places, better than the server's ten highest-rated players on every statistic (2.34, 59%). TavernBench, 960 tasks in 40 scenarios, shows that the system the agents built is the strongest player: TavernAgent answers 91% of the play tasks against at most 63% for frontier LLMs, and decides in 0.03 s where they need up to 26 s; within the second a live turn allows, it keeps 90% and no LLM more than 26%. LLM agents build far better than they play. We will open-source TavernBench and TavernGym at https://github.com/MCG-NJU/AutoTavern.

Abstract: On-policy distillation aligns teacher feedback with a student's current behavior, yet acquiring useful supervision for multi-turn agents requires identifying both where guidance is needed and whether the student can sustain the interaction afterward. We introduce Frontier-OPD, which addresses this problem through the student's learning frontier: decision situations that remain challenging but admit useful autonomous continuation after brief teacher guidance. To operationalize this frontier, Frontier-OPD organizes on-policy experience into a shared-state trajectory graph, where trajectory-balanced aggregation of teacher-student disagreement identifies distinct intervention candidates while preserving the histories needed for replay. Short teacher bridges are then assessed by comparing agreement over the bridge and the ensuing student continuation with an equally long segment of the original trajectory, providing a behavioral proxy for guidance that the current student can build on. Accepted teacher responses augment ordinary distillation data, and fresh rollouts renew the frontier estimate as the policy changes. Experiments on ALFWorld, WebShop, and ScienceWorld report higher mean autonomous performance than the compared distillation baselines across the evaluated student scales and teacher configurations, with ablations indicating complementary effects of candidate selection and guidance filtering. In one ALFWorld timing setting, the additional collection cost is offset by a shorter time to a target success rate.

Abstract: Building virtual embodied agents that operate in open-ended games like human players is a long-standing goal requiring both long-horizon autonomy and reliable native-interface control. As a procedurally generated open world, Minecraft provides a natural testbed for studying the integration of these capabilities. Existing agents typically provide either native visuomotor control without long-horizon autonomy or high-level planning executed through environment APIs or predefined skills. To bridge this gap, we propose HiSTEVE, a hierarchical VLA agent with a capability-aware agent harness. At the orchestration level, we equip the harness with structured memory for tracking progress and a VLM planner for selecting subtasks within the VLA's capabilities. To reliably determine subtask completion, we further introduce a policy-coupled task completion state detector that conditions on VLA policy context to signal candidate completion. At the control level, a single VLA executes all subtasks with vision-only perception and native keyboard-and-mouse actions. To mitigate interference between heterogeneous GUI and non-GUI interactions, we introduce regime-specific RMSNorm branches and separate mouse-movement token embeddings. We also train the VLA through world knowledge injection, imitation learning, and reinforcement learning with environment feedback to improve reliability and recovery. HiSTEVE achieves state-of-the-art atomic-task performance, scoring 85.0% on MineStudio Bench and a 37.4% average success rate on OpenHA, while also completing long-horizon tasks such as crafting an iron pickaxe from scratch like human players.

Abstract: LLM agents are flexible in language interaction, but using them as long-horizon planners is costly and prone to hallucinated or invalid actions. This paper targets a different role for LLMs: constrained bridges between language and symbolic planning. We propose MemPlan, a memory-conditioned PDDL planning framework for partially observable interactive text environments. MemPlan keeps the PDDL domain fixed and reconstructs the PDDL problem at each step from parsed observations, candidate hypotheses, and interaction memory. Positive graph memory assigns relevance-based action costs to candidate hypotheses for missing objects, locations, or conditions, guiding the planner toward more plausible checks. Negative memory records execution-refuted assumptions and compiles them into exclusion constraints, preventing repeated attempts at candidates already contradicted by feedback. A unified schema-conditioned language--symbol interface, trained offline with supervised fine-tuning, connects text environments to the planning loop through observation-to-fact parsing and symbolic-action-to-command grounding. Experiments on TextWorld, ALFWorld, and Robotouille show that MemPlan improves task completion and step efficiency over language-only and PDDL-aware baselines, while substantially reducing online token cost compared with LLM-planning baselines such as ReAct. The same fine-tuned interface is also reused across TextWorld task types without task-specific fine-tuning. Our code and data are available at: https://github.com/tewiSong/MemPlan.

Abstract: Large language model agents are often given an inference budget and asked to maximise success within it. A multi-turn agent has to allocate this budget across many environment interactions. An external harness can enforce the budget, but it remains unclear whether the agent itself can learn when and how much to spend. We ask whether agents can learn this allocation across dynamically specified budgets and which rewards best teach this behaviour. We adapt existing length penalties to budget constraints spanning multiple resource types, introduce new reward formulations, and compare eleven objectives in a shared training pipeline. We train Qwen3.5 models at 4B and 9B on AppWorld with budgets over tokens, API calls, and environment steps, and on the BabyAI gridworld with turn budgets. We train a separate policy for each environment and resource type, with each policy trained across budget values. Training with rewards that penalise overspending substantially improves within-budget success and natural compliance over untrained models and over training on task success alone, with agents learning to adapt their resource use to stated budgets and task difficulty. These gains extend to unseen budgets and, zero-shot, to a shifted task distribution and a new benchmark. Reward objectives trade off success within budget and natural compliance. A sigmoid penalty for overspending most often achieves the highest within-budget success at an individual budget, while a fixed negative reward for overspending generally yields the strongest natural compliance. Combining penalties for overspend frequency and magnitude gives the smallest average gap to the best method at each budget. Overall, we show that a single agent can learn to allocate resources across turns and adapt its behaviour to user-specified budgets using budget feedback without external budget-allocation controllers.

Abstract: Video world models turn actions and camera motion into long visual rollouts, making video a natural interface for interactive simulation. Such interaction requires responsive generation, yet coherent rollouts must retain visual memory of what has already been generated. Chunked video diffusion models usually keep this context in growing KV caches, so later chunks become progressively slower. To address this bottleneck, we introduce Velox, a training-free attention router for region-level context selection using frustum slab correspondences from streaming geometry. Velox represents historical blocks as camera frustum slabs with depth intervals and obtains current-block depth through recency-priority multiwarp. For well-covered regions, frustum slab overlap softly modulates each layer's model-derived block importance. Uncovered regions retain the model-derived route without geometric bias. The fixed token-block granularity matches FlashAttention-style tiled execution and reduces the attention term that grows with rollout length. Frustum slab construction runs asynchronously with denoising on the same H200. At chunk 48, Velox accelerates attention by $4.36\times$ and denoising by $2.43\times$. On the full 158-case WBench navigation split, it remains within 0.15 API-free average points of dense LingBot-Fast-Cam. Please see our project page at https://velox-world.com/velox/ for more details.

Abstract: Incremental learning (IL) aims to continually acquire new knowledge (plasticity) while retaining previously learned information (stability). However, striking a balance between plasticity and stability remains a significant challenge for intelligent systems. The human brain achieves exceptional balance, owing to various memory units that collaboratively encode and store information. Inspired by human memory mechanisms, this paper introduces an IL framework with knowledge enCoding, Consolidation, and Integration (Co2In). Co2In is designed as a bi-level memory architecture: a working memory for adaptive knowledge acquisition and a long-term memory dedicated to the persistent retention of information. The two memory modules work cooperatively during IL. The working memory first learns from new data to encode knowledge into parameters. Subsequently, Co2In performs a consolidation process to identify underlying patterns in the learned parameters and re-express them into a compact knowledge representation. Next, the knowledge representation and the identified patterns are transformed into separate network layers and integrated into the long-term memory. These designs empower Co2In to accumulate knowledge with high plasticity and stability. We evaluate Co2In on CIFAR-10, CIFAR-100, and Tiny-ImageNet under exemplar-free Class-IL and Task-IL settings. Experimental results show that Co2In achieves state-of-the-art performance with efficient memory consumption.

Abstract: Transformers can implement reinforcement-learning algorithms in context, yet how off-policy control emerges within Transformers remains poorly understood. We study Q-learning, a foundational model-free control algorithm that combines off-policy bootstrapping with greedy policy improvement. We prove that a two-layer attention-only Transformer can implement a finite-temperature form of Q-learning, where the temperature $τ>0$ controls the approximation to greedy action selection, recovering window Q-learning as $τ\to0$. We train the structured Transformer on randomly generated MDP tasks using output-only supervision. Experiments show that the theoretically predicted parameter structure emerges during training, including the internal greedy action-selection mechanism, and that the trained Transformer achieves performance comparable to Q-learning on unseen tasks. Our theoretical analysis further shows that the equivalent realizations form a solution manifold and that, under suitable task-diversity conditions, the training iterates converge locally to this manifold at an exponential rate. For the random-task model analyzed here, the local worst-case number of gradient steps needed to reduce the distance to the manifold by a factor $ε$ is $Θ\left(τ^{-5}\log(1/\varepsilon)\right)$, quantifying the increasing optimization cost as the learned update approaches window Q-learning.

Abstract: GPU architecture research requires representative benchmarks that can execute on new designs. However, many vendor-tailored workloads and library kernels cannot run directly on research architectures because their implementations are closed-source or rely on programming models unsupported by the target architecture. Migrating existing kernels requires manual effort to adapt their implementations and validate their behavior. This creates a need to automate workload migration to target platforms. We present VortexSynth, a framework that uses profile-guided benchmark reconstruction to construct GPU benchmarks from reference workloads. Our workflow selects hot kernels through profiling, synthesizes implementations in the target programming language, verifies output equivalence, and measures profile differences using a translated proxy implementation. The resulting profile signature gap is supplied to an agent for iterative refinement. We evaluate the method on Vortex, an open-ISA GPGPU, using robotics as the application domain. All active kernel pairs pass both output-equivalence checks, and the case studies show that iterative refinement can reduce profile distance.

Abstract: Interactive virtual worlds allow agents to learn through exploration and interaction. What agents can learn is bounded by the environments they practice in, which must be faithful, with consistent state, rules, and dynamics, and realistic, with observations that follow the real-world visual distributions. Achieving both across diverse worlds remains a bottleneck. We introduce AgentGarten, a framework that couples simulators and game engines with a shared neural renderer to build real-time interactive environments. Its simulation backends maintain persistent world state and execute program-defined interaction rules, while the renderer generates visual observations from structured conditions exported through a common interface. To build the neural renderer, we adapt a pretrained video model to geometry conditions, distill it with our proposed Adversarial Forcing, and optimize inference for real-time interaction. Adversarial Forcing makes history prefilling differentiable through exact replay, so that losses on later predictions update how the renderer encodes prior observations, and adds real-data adversarial supervision to improve its visual quality. In AgentGarten, agents perceive the world through visual observations, interact with it in real time, and improve by distilling each round of experience into playbooks that subsequent agents inherit and refine. Our empirical study demonstrates a substantial gain in learning efficiency, with agents learning from just 4 rounds compared with millions for a conventional reinforcement learning counterpart. As new worlds can be written as code and rendered through the same interface, environments can scale in both number and difficulty alongside their agents, a step toward agents that keep evolving through interactive experience.

Abstract: Foundation models supply reinforcement learning (RL) with priors that mitigate its longstanding weaknesses in sample efficiency and transfer, but their token-by-token generation makes queries sequential and costly. Jev, a recently released decision model, generates nothing and returns calibrated, typed answers in a single forward pass. Existing work studies foundation models in RL either as models to be trained or as generators to be prompted, and Jev belongs to neither category, having so far served only as a black box in single domains. How well such a model decides on its own in RL environments, and how it can improve RL as a component of training, therefore remain unaddressed. To this end, in this paper we first examine the requirements that the objects of an RL system place on the answers they consume, and establish that Jev can fulfill all of them except the cardinal use of a value function. The remaining objects form positions that admit several roles each. We then construct algorithms that employ Jev at three of these positions, as a reference policy, an exploration judge, and a replay rater, to improve sample efficiency, exploration, and learning performance. Across nine MiniGrid tasks and three Atari games, training with Jev outperforms a standard RL learner, including where the learner makes no progress alone, while the model itself remains untrained. To our knowledge, we present the first use of Jev within the RL learning process and establish a frozen decision model as a usable component of RL training, inviting further exploration of how Jev and other advanced decision models can improve RL.

Abstract: Memory-augmented reinforcement learning strengthens LLM agents' ability to solve complex long-horizon tasks. Skills are one such form of memory, pairing instructions with an applicability condition over task types. However, retaining every skill indiscriminately as the policy improves lets obsolete or harmful entries accumulate and mislead the agent. We propose SkillForge, an agentic RL method that compiles and evolves the skill library through a fitness-driven skill lifecycle of trial, active, stable, and retired states, so that the skills and the model co-evolve throughout training. A pre-RL evaluation phase first uses the base model's own rollouts to pre-retire low-fitness skills, yielding a filtered library that then seeds supervised fine-tuning. Reinforcement learning takes over from this checkpoint, and at each iteration selective retirement, stabilization, and LLM-guided mutation continue to forge the skill library alongside policy optimization. Across multiple interactive agent benchmarks, SkillForge achieves the highest aggregate success rate, delivering up to 7.8% relative improvement over the strongest baseline while keeping the skill library compact throughout training. We introduce SkillFurnace, a dataset of 5k+ annotated records bundling retirement-filtered SFT trajectories, evolved skill libraries with fitness annotations, and retirement events with human-annotated failure categories to support research on skill quality and lifecycle management.

Abstract: Frontier LLM agents have been shown to be capable of solving increasingly complex tasks for which humans have measurable solutions. This begs the pertinent question of whether LLM agents can go beyond what humans have already solved. The ability to develop sophisticated strategies to tackle consequential problems becomes paramount as well-trodden, human-developed solutions become insufficient for problems for which we lack context or enough training data. We study agents' capability of such strategy formation through the communal practice of video game speedrunning. In speedrunning, practitioners compete to find the fastest way to complete a video game under certain conditions, and in so doing uncovering interesting unorthodox play styles that require a thorough understanding and mastery of the underlying game mechanics. We introduce SPEEDRUNBENCH, a benchmark that evaluates frontier LLM agents across 9 different games. To perform well in this benchmark, agents must repeatedly improve their strategy, reflect on their performance, exploit their gained knowledge, and reason across a long-horizon of actions to improve on an increasingly difficult problem: being faster than themselves and everyone else. Our experiments show that while frontier agents approach human world records in simple platformer games, they remain behind human performance on longer, more complex games under practical budgets. These results suggest that SPEEDRUNBENCH is a useful testbed for studying agents' strategy formation capabilities as well as being a saturation-resistant evaluation measure, as there is almost always a faster completion time waiting to be discovered.

Abstract: Where generalization capacity--the ability to extract context-invariant relational structures--first emerges remains a central question in AI and neuroscience. The foundation for this capacity lies upstream of the hippocampus, within the entorhinal cortex, where parallel pathways dissociate relational structure in the medial entorhinal cortex (MEC) from sensory content in the lateral entorhinal cortex. However, as Eichenbaum argued, such factorization likely originates earlier, driven by the segregation of the dorsal ('where') and ventral ('what') visual streams. Supporting this, grid-like firing patterns--a signature of MEC (context-invariant codes)--also appear in preceding neocortical regions along the dorsal pathway. Yet, how such representations are computationally formed along upstream pathways remains unknown. To investigate this in silico, we developed ReGraph, a recurrent dual-stream graph model with biological inductive biases, including retina-driven stream-specialized encoding, dorsal-to-ventral modulation, and dynamic lateral connectivity. Trained on the action benchmark Something-Something V2, ReGraph revealed a pathway-specific emergence of relational mapping: context-invariant codes and grid-like spatial bases uniquely co-emerged along the extended dorsal stream. In contrast, their absence in single-stream, unmodulated variants, and standard baselines implies that these inductive biases are prerequisites for relational structures. Crucially, our post-hoc analyses demonstrated that these grid-like bases serve as reusable routing templates for information processing via lateral connectivity. Together, our findings provide a computational account that generalization may not be a faculty that emerges abruptly within a dedicated region, but a property that already takes shape as sensory information is parsed into factorized streams of hierarchical visual processing.

Abstract: Understanding a nonlinear dynamical system from time series requires not only reproducing its trajectories, but also identifying a simple representation that preserves its essential dynamical structure. Almost-linear recurrent neural networks (AL-RNNs) are piecewise-linear RNNs in which only a subset of units use ReLU nonlinearities, so that nonlinear capacity is explicitly controlled by the number of ReLU units. Their activation patterns define linear regions, represented as symbols, whose observed transitions form a symbolic transition graph. However, directly training AL-RNNs with few ReLU units to realize minimal dynamical representations can be unreliable. We ask whether an AL-RNN with more ReLU units can instead be trained first and systematically reduced to a minimal dynamical representation. We introduce a fixed-point-guided hierarchical reduction procedure that progressively linearizes selected ReLU units, merging neighboring linear regions and graph nodes while preserving distinct symbols containing fixed points (FPs). The resulting reduction tree defines a hierarchy of progressively simpler candidates. Each reduced candidate is initialized from the parent parameters and retrained under guidance from the parent dynamics. We also prove that reproducing $Q$ distinct fixed points requires at least $Q$ FP-containing symbols, providing a certificate of symbol-level minimality when this bound is attained. On the 3-scroll Chua system, direct training with the theoretical minimum of three ReLU units achieves high-fidelity minimal realizations in only 20% of seeds, whereas our learn-reduce-retrain strategy increases the seed-macro success rate to approximately 71% at the same final nonlinear capacity. These results show that redundant nonlinear capacity can serve as a scaffold for discovering and realizing minimal dynamical representations.

Abstract: Many visual tracking methods use rejection mechanisms to suppress unreliable predictions. However, these mechanisms can also reject correctly localized candidates, leaving useful information unused. We investigate how to identify and recover these candidates while preserving native accepted outputs and candidate coordinates. To this end, we propose P-SRM (Post-rejection Selective Recovery Method), which combines spatial responses, past accepted states, and native decision margins to reassess candidates and selectively restore reliable predictions. We evaluate P-SRM on six trackers and four datasets spanning category-specific, point, and generic object tracking. Across all nine configurations, P-SRM improves rejected-candidate ranking and overall tracking performance. These results show that post-rejection verification can identify and recover useful predictions discarded by native rejection, demonstrating the value of reusing rejected information. Project repository: https://github.com/PalestyHR/P-SRM.

Abstract: An agent that remembers what it did on a web page must decide when two pages count as the same. Memories built on observation similarity merge pages that look alike but behave differently, and GUIs are full of such pages: two tabs of one widget or two rows of one menu answer the same click differently. We define the merge rule as an action-conditioned bisimulation over the empirical predictive state graph a frozen agent fills as it acts. Two states merge only when their shared actions lead to agreeing outcomes and successor blocks under an affordance label. Observation similarity never enters the rule, and nothing is trained. It replaces the merge rule of an existing outcome-value memory, so a closed-loop comparison isolates it. On MiniWoB++ it raises success rate over a memoryless agent, while a control taking identical exploratory detours, the prior successor-representation merge, and the same criterion without action conditioning change nothing.

Abstract: Autonomous-driving question answering requires reasoning over structured scene information, yet existing vision-language approaches largely delegate heterogeneous reasoning operations to a single neural inference process. We argue that this uniform strategy overlooks a fundamental distinction: some queries admit exact symbolic solutions, while others require semantic interpretation. We introduce a query-adaptive neuro-symbolic reasoning framework that explicitly allocates computation according to the nature of the query. At its core is a hierarchical Spatiotemporal Scene Graph (STSG) that separates persistent object identities from frame-specific states and represents spatial relations and temporal transitions as explicit directed structures. Given a query, a symbolic executor first attempts to resolve it through exact graph operations; only when symbolic execution abstains is an LLM invoked for semantic reasoning. For these unresolved queries, query-conditioned graph retrieval and evidence filtering preserve relation direction, temporal locality, and object semantics, providing the LLM with compact and verified task-relevant evidence. This design shifts the role of the LLM from a universal reasoning engine to a targeted semantic reasoner, while allowing deterministic computation to be handled exactly and efficiently. We evaluate the framework on 5,916 NuScenes-QA questions across all ten scenes of nuScenes v1.0-mini under an oracle-perception setting. The complete system achieves 80.63 percent overall accuracy with GPT-5.4-mini, improving over the corresponding LLM-only configuration by 5.48 percentage points; with DeepSeek-V4-Flash, the improvement reaches 6.64 points. The largest gains occur on counting questions, with improvements of 10.20 and 12.61 points, respectively. These results show that selective reasoning improves both accuracy and inference efficiency.

Abstract: State tracking requires composing a sequence of updates, but accuracy alone does not reveal what a model has learned. We study neural networks trained to predict the running product of group elements. We identify quotient solutions in Transformers, where models recover the quotient class while predicting nearly uniformly among its members. The reciprocal of class size predicts partial accuracy without a fitted parameter, extending parity-based accounts to non-parity quotients. Our baseline Transformers' predictions change little under prefix reordering beyond the exact-tracking frontier. We prove that, for finite groups under uniform i.i.d. full-group inputs, optimal order-blind exact accuracy converges to the reciprocal of abelianization class size as prefix length grows, consistent with the observed abelianization plateaus. Sequential updates permit more: any partition into right cosets of a subgroup, normal or not, survives sequential updates. In our census of standard Transformers, every recovered coset partition comes from a normal subgroup, whereas parameter-matched recurrent networks pass through both normal and non-normal right-coset stages during training. On $A_5$, we identify low-dimensional subspaces of the recurrent state that encode non-normal cosets. In the three-dimensional cases, coset mean vectors form approximate dodecahedra, and swapping the state components in these subspaces transfers the donor's coset state through a shared input suffix. Our results connect partial accuracy, learning stages, and internal computation through the subgroup cosets that models learn to track.

Abstract: Vector Symbolic Algebras project data structures into a hyperdimensional vector space through the application of their vector algebras to randomly generated atomic vector symbols and fractional power encodings of real-valued data. Embedding unstructured data remains an open question. We present \textit{Clifford-VAE}, a variational autoencoder that learns to project data onto a Clifford torus in arbitrary dimensions. Experiments using the MNIST, FashionMNIST, and CIFAR-10 datasets demonstrate that Clifford-VAE produces representations that are competitive with those produced by Gaussian and Hyperspherical VAEs for semi-supervised classification tasks while outperforming Gaussian and Hyperspherical counterparts in the VSA benchmark tests of self-binding and unbinding, role-filler recovery, and bundle capacity. Clifford-VAE provides a principled technique for grounding perceptual data into a symbolic reasoning framework, providing a new approach to a long-standing problem in the VSA literature.

Abstract: Long horizon agents accumulate growing interaction histories that increase context and inference costs. We find that geometric redundancy alone is an insufficient criterion for safe compression. Although agent histories exhibit strong low dimensional structure, similar global geometry can preserve very different amounts of task evidence. At identical retained block counts, evidence aware selection raises next action Top 3 retention from 0.31 to 0.69, while centroid similarity remains 0.98. Controlled replacement further shows that action related information can be substantially altered while global geometric measures remain nearly unchanged. Motivated by this gap between geometry and evidence, we introduce Geometry Guided Evidence Preserving Memory (GEM), a training free compressor that protects task and execution evidence before using geometric residuals to complete coverage. GEM reduces mean combined token usage from 2.69M to 2.11M per task, a 21.4% reduction, while maintaining comparable task reward. Our results show that efficient agent history compression should optimize for preserved task evidence rather than geometric coverage alone.

Abstract: As Large Language Model (LLM) agents are applied in continuously interactive environments, driving the evolution of their own capabilities becomes a core problem for achieving long-term autonomy. Currently, environmental knowledge is typically treated as an external fixed input rather than as part of the agent's ongoing evolution. Reinforcement learning methods usually optimize policies through environmental interaction but tend to adapt only to fixed task distributions or single environments. This paper proposes TTSE (Two-Track Self-Evolution), a dual-track online self-evolution framework that separates evolving knowledge into FACT (environmental facts, whose reliability is continuously verified through interaction evidence) and TIP (task-conditioned implementation procedures). From a decision-theoretic perspective, we decompose the agent's excess risk into environment-representation regret and conditional-execution regret, characterize the conditions under which environment-conditioned policies strictly outperform condition-agnostic policies, and bound the downstream risk in terms of FACT identification error and cross-condition mismatch cost. In practice, TTSE's ablation experiments on GDPevo validate the advantage of dual-track evolution. On the classic agent task benchmarks ALFWorld and ScienceWorld, TTSE further demonstrates superior task adaptation. Moreover, TTSE is broadly compatible with existing skill self-evolution methods; combined with the Bayesian-Agent algorithm, a single-track ablation validates the dual-track advantage, substantially improving the aggregate score across the five major domains of SOPBench over three independent repetitions. Finally, on the real end-to-end task benchmark PinchBench, TTSE is integrated into a general agent framework via retrieval-based injection and stably outperforms the baseline across three independent runs.

Abstract: In recent years, pre-training has become fundamental to learning effective video representations, enabling strong transfer to downstream tasks. A popular framework in pre-training involves aligning features of a video encoder with that of another modality, for example, language or audio. We introduce Video-STLayout pre-training, a novel strategy for obtaining rich video representations informed by spatio-temporal layout of object bounding boxes. Object layouts can easily be obtained by applying an off-the-shelf object detector on the video frames. Our method uses a contrastive loss to align video features with the layout features from a trained layout encoder. We show the effectiveness of our approach in the task of activity recognition in complex scenes.

Abstract: Language models compute over tokens: language is their input, their output, and increasingly their internal representation. Whether language should keep all of these positions depends on what language does to the system that uses it. The one system with a century of data on that question is the human. We review what language does to human perception, the brain, and thought, and read the same evidence against multimodal models and language models. Throughout, we treat language as a compressor that runs on a shared codebook: a word is an index, the content is in the receiver, and a community maintains the codebook. In humans the compression is measurable, learning the codebook reorganizes the senses, and thought survives the loss of language. We then measure the rule that models apply when two cues disagree, with cue-conflict experiments on six vision-language models and two robot policies. Surviving cues are weighted in the order their reliabilities prescribe, at 11 to 82\% of the ideal observer's slope, and many answers copy the text. One policy family drops a cue that adds no information beyond the others rather than down-weighting it, another keeps it at a weight that fails when the cues conflict, and a visual cue that identifies the task in every training frame is never learned, because the language pathway already fits the data. Language models are the best current models of the human language network, and they have entered the human speech community, shifting word frequencies while alignment narrows their conceptual diversity. We close with seven implications for token-based systems. Language belongs at a model's boundary and in the shared codebook, as in the brain, not as its internal representation; the price of leaving the codebook inside is auditability.