Papers

LLM coding agents now generate complex programs at a scale that makes thorough human review increasingly difficult, raising the risk of safety and security failures. Common approaches, including fuzz testing, static analysis, and LLM-as-a-Verifier, can detect many failures but struggle to cover all possible edge cases. Formal verification addresses this by providing machine-checkable guarantees over specified properties, but traditionally demands substantial manual specification and proof engineering. We introduce a unified multi-agent framework, MAGS, that generates executable programs with formal safety guarantees, using Dafny as a verification-aware intermediate representation where safety properties can be mechanically checked. MAGS formalizes and freezes human-audited APIs and safety requirements, translates generated code into Dafny, repairs violations using verifier feedback, and compiles verified programs back into executable code. We evaluate MAGS on 100 CUDA kernels, 100 terminal scripts, and 20 robotic-arm tasks. Across all 220 examples, it achieves a 100% success rate in producing programs with non-trivial safety guarantees against frozen specifications. Independent safety and functional evaluations further show strong performance across all three domains, while revealing failures when the auto-formalized semantics do not fully capture the target behavior.

Agents are increasingly asked to design hardware, and increasingly reported to succeed. Such reports establish that a design improved; they cannot establish why. An agent that improves an accelerator may be reasoning about the machine, or may be searching competently over knobs whose meaning it never recovers – and only the first transfers to the next architecture. Existing evaluations cannot tell the two apart, because they vary the agent while holding the framing of the problem fixed. We do the opposite. AutoTuring hands the same agent the same 15-dimensional accelerator space twice: once as named architectural knobs with simulator counters, once as anonymous variables on [0,1], with the evaluator, the legal space and the reachable optima held identical, so that the only thing that varies is whether the problem means anything. The gap between the two is the measurement. On a nine-kernel FP16 GEMM basket, meaning pays: the architect beats a modeled H200 by 5.4% and its blind counterpart by 12.3% on average, with 70.1% fewer simulator calls. It does not pay uniquely: a critic loop recovers most of that gap for the blind agent and buys the architect nothing, so architectural knowledge and structured critique behave as substitutes rather than as complements. We report these as preliminary findings – five to six runs per condition on a single modeled accelerator – and take the comparison itself, not the accelerator, to be the contribution.

Scaling deep learning faces critical bottlenecks: data exhaustion, exponential training costs, and resource concentration. Model merging combines pre-trained checkpoints without gradient descent, offering orders-of-magnitude savings versus retraining. Combining independently trained vision models is difficult when their architectures and parameter shapes differ. Existing weight-space merging methods generally assume aligned, shape-compatible checkpoints, whereas a Vision Transformer (ViT) and a state-space model (SSM) implement token mixing with different operators. We study a hybrid Heterogeneous merging setting that retains both architectures while aligning parameter groups by semantic role. Our proposed Riemannian–Lorentz Parameter Fusion (RLPF) method projects aligned groups to common coordinates, lifts selected coordinates to the Lorentz hyperboloid model of hyperbolic space, computes a regularized geodesic barycenter, and decodes the result into the two branches. A learned gate then combines branch logits for each input. Component groups use fixed curvature values, with normalization parameters treated as Euclidean. In the results available in this manuscript, the fine-tuned system obtains 82.37% on CIFAR-10, 75.04% on Oxford-IIIT Pet, and 78.58% top-1 accuracy on ImageNet-1K; the corresponding best-parent accuracies are 76.54%, 71.42%, and 76.42%. On ImageNet-1K, the reported pre-fine-tuning initialization reaches 77.80%. These results support further study of geometry-aware heterogeneous fusion, but not a training-free single-checkpoint merge: RLPF is a two-branch hybrid whose gate and reported final models are trained.

Recovering numerical series from line plots requires accurate axis calibration and reliable curve extraction. We present LinePilot Digitizer (LinePilot), which combines continuous color-based curve recovery with three calibration modes: LinePilot (standard), LinePilot (enhanced), and LinePilot (OCR). We also introduce DigitizerBench, the first dedicated benchmark for systematically evaluating digitizer performance, using an orthogonal design spanning signal, rendering, and plot-structure factors with complementary automatic and human-guided evaluations. We evaluate performance using failure-penalized capped normalized root-mean-square error (FPC-NRMSE), which assigns unit loss to missing, unusable, or catastrophically inaccurate outputs. On DigitizerBench-Full, LinePilot (OCR) achieves the lowest mean FPC-NRMSE (0.672) and highest trusted usability (38.2%) among the tested automatic pipelines. On DigitizerBench-Lite, LinePilot (enhanced) achieves the lowest mean FPC-NRMSE (0.081), 100% output success, and highest trusted usability (93.3%). The orthogonal benchmark design further enables factor analysis to identify the factors that most significantly affect digitizer performance. Together, the three calibration modes provide a practical trade-off between automation, user control, and accuracy within a shared curve-recovery workflow.

We introduce a new method to guide flow matching models. Our approach, which we call probe guidance, uses the frozen internal states of an existing diffusion model to construct a guidance signal. This works using a similar principle as autoguidance, but eliminates the need for an additional forward pass at inference time and provides a reliable path to ensure that the weak and strong model share similar dynamics. We apply and benchmark this method on continuous diffusion language models, where probe guidance sets a new state-of-the-art performance on unconditional generation. When applied to a 1.7B diffusion language model, probe guidance consistently improves on multiple choice question answering benchmarks. Using our probes, we study the traditional autoguidance setting where the strong model is a weak checkpoint, and find that the weak model must come from a low-entropy region of training. These findings both provide a practical way to improve diffusion language models and shed light on the actual mechanism behind autoguidance, which is currently poorly understood.

Automated action quality assessment (AQA) in Olympic sports remains a challenging task due to the complexity of human motion and the subjectivity inherent in expert judging. This work evaluates the capability of open-source Vision-Language Models (VLMs) to perform zero-shot action quality assessment on Olympic diving videos using the AQA-7 benchmark dataset. In this regard, a regression-based framework is pro-posed to leverage both the semantic reasoning and phase-level sub-scores generated by the VLMs, combining TF-IDF vectorization, dimensionality reduction, and ensemble learning to predict final competition scores. Experimental results show that standalone VLMs achieve moderate Spearman correlations below 0.32, while the proposed ensemble regression framework substantially improves performance in the reported evaluation, reaching a Spearman correlation of 0.67 with a four-model configuration. Textual reasoning features con-sistently outperformed raw numerical sub-scores, highlighting the richness of VLM-generated explanations for action quality analysis. These findings suggest that VLMs hold strong potential as assistive tools for explainable and semi-automated sports performance evaluation. The code is publicly available on GitHub https://github.com/hvelesaca/olympic diving judge vlm

Robotic additive manufacturing (AM) extends material-extrusion printing beyond gantry kinematics but makes process planning robot-dependent. A slicer-generated plan that appears favorable in part coordinates can become infeasible or robotically unfavorable on a manipulator because slicer-process decisions and part orientation determine the generated path, while part orientation and workspace placement affect its kinematic realization. Existing AM tools, large language model (LLM)-based decision-support methods, and digital-shadow systems do not provide integrated pre-execution evaluation of these coupled decisions. This paper presents agentic robotic additive manufacturing (A-RAM), an agent-specialist-tool framework that converts user intent and a part file into traceable, execution-ready plans. The LLM interprets manufacturing objectives and constraints, identifies prescribed and searchable planning variables, and encodes this reasoning in a schema-constrained request; a deterministic Planning Agent instantiates the corresponding search workflow, while domain tools compute quantitative evidence for slicing, placement, inverse kinematics, trajectory timing, Joint-6 jerk, and extrusion. The framework is evaluated on a six-axis robotic-arm AM cell through three case studies covering expert-specified planning, goal-only planning, objective-dependent infill screening, and geometry-dependent orientation-placement selection. Across the evaluated candidate sets, selected plans achieve up to 53.5% lower maximum Joint-6 jerk and 48.3% lower mean absolute Joint-6 jerk than the least favorable valid candidates, while objective-specific infill screening yields motion-plan completion times up to 40.1% shorter and extrusion paths up to 12.7% shorter than the corresponding least favorable screened patterns.

Safety tuning pipelines judge only the final answer, which makes it difficult to distinguish robust refusal from two undesirable shortcuts: blanket refusal on benign requests and polished but unfaithful safety rationales that do not actually constrain the answer. We propose AUDITPLAN, a single-model plan-then-answer approach where the model first emits a compact structured safety plan and then answers conditioned on it. The plan records a threat label, intended action, and explicit constraints, enabling machine-checkable auditing while remaining hidden from users at deployment. We train this behavior with supervised fine-tuning followed by reinforcement learning with FAITHGATE, a reward-gating objective that grants answer reward only when the safety plan is correct. This discourages safe-looking but unfaithful behavior and promotes tighter plan-answer coupling. Across Qwen backbones, AUDITPLAN improves both robustness and auditability: on Qwen2.5-3B-Instruct, FAITHGATE reduces ASR from 24.0% to 11.6%, LSR from 1.0% to 0.36%, and over-refusal from 11.0% to 2.0%, outperforming answer-only RL, free-form explanation, and weighted-sum structured rewards. Similar trends hold for Qwen2.5-1.5B-Instruct. Larger-model confirmation runs on Qwen-3-4B-Instruct and Qwen2.5-7B-Instruct preserve the same trend suggesting that explicit internal commitments can make safety alignment more faithful, robust, and auditable.

Large language models (LLMs) provide a flexible interface for long-horizon robot planning, but generated plans often fail to respect embodiment constraints, recover from planning errors, or reason effectively under partial observability. We present GAVEL, a framework for verifying and repairing long-horizon LLM planning built around an explicit graph world model. The graph represents relevant object-relations, action pre-conditions and effects, and probabilistic beliefs over unobserved object locations. This model can predict the consequences of LLM-generated actions before execution, detect violations, and repair those whose corrections follow directly from the world model. This method also reserves LLM replanning solely for errors requiring semantic reasoning. For multi-task instructions, GAVEL reasons over distributions of possible object locations to reorder remaining subtasks and minimize expected search cost. We evaluate GAVEL on BEHAVIOR-1K across 100 single long-horizon tasks and 500 multi-task instructions. With Qwen3-8B, GAVEL improves single-task success from 41.2% to 91.8% and multi-task success from 19.9% to 92.6%. Distributional belief reasoning also reduces travel distance by approximately 5.4% compared with a static variant. These improvements show that an explicit graph world model harness can substantially improve the reliability and efficiency of long-horizon embodied planning across compact and frontier hosted LLM capabilities.

Large Language Models (LLMs) have made remarkable progress in the processing and modeling of many languages. Yet, unlike human multilinguals, they exhibit surprisingly limited cross-lingual knowledge transfer. While this limitation is well documented, its origins during multilingual training remain unclear. We pretrain 360M- and 7B-parameter LLMs and show that poor cross-lingual knowledge generalization emerges during pretraining and persists under standard interventions. To isolate its cause, we employ a controlled bilingual pretraining setting using two copies of the same language, sharing identical text and token segmentation, but mapped to disjoint token spaces. We find that disjoint tokens alone are enough to induce knowledge compartmentalization, even between identical copies of the same language, establishing disjoint token spaces as a fundamental barrier to cross-lingual knowledge generalization. Guided by this understanding, we suggest mapping languages into a shared token space by simple word-wise translation and find it substantially improves cross-lingual knowledge generalization, recovering up to 12.6% of native-language learning efficiency — 14$\times$ the baseline.

Clinical decision-making for coronary intervention relies mainly on angiography and fractional flow reserve (FFR). However, angiography is two-dimensional and lacks depth information for 3D lesion characterization, while FFR provides only a single functional index, offering limited hemodynamic insight. Among existing methods, numerical analysis is computationally expensive, whereas learning-based approaches require extensive supervision and often lack physical consistency. To address these limitations, we propose physics-informed hemodynamic modeling, an integrated deep learning framework for 3D coronary blood flow analysis from dual-view angiography. First, an attention-enhanced CNN reconstructs coronary geometry from angiography. The resulting point clouds are then mapped to a reference domain and Fourier-encoded for joint representation. A decoupled network separately predicts velocity and pressure fields, with embedded physical priors enabling efficient transfer across physiological conditions. Across 32 clinical patients evaluated under four flow conditions, the trans-stenotic pressure-drop mean absolute percentage error was 2.02%, while the velocity and pressure relative-L2 errors were 0.054 and 0.023, respectively. Validation against hospital-measured FFR further achieved 93.8% diagnostic accuracy (30/32; exact 95% CI, 79.2%-99.2%). The framework also supports illustrative revascularization comparisons and sparse-data assimilation, with the full angiography-to-hemodynamics pipeline completed within 20 minutes per patient.

Conversational LLM agents increasingly rely on Web search, yet the end-to-end lifecycle of agentic search remains poorly understood. We present the first study of Web search across four major conversational platforms (ChatGPT, Claude, Grok, and DeepSeek), combining real-world user interactions (invivo) with controlled experiments using the same platform's models by their APIs (invitro). We investigate the quality of agentic decisions to invoke Web search, their strategies to formulate queries, the potential domain preferences in the search results they receive, and the choices they make when transforming search results into grounded responses. We find that Web-search decisions vary substantially across platforms and models, while more frequent Web-search invocation does not necessarily yield better response quality. We further show that conversational agents employ different complex querying strategies and that platform specific search engines return search results from their preferred domains. Finally, although responses are largely grounded in search results, some claims rely on uncited search results, raising concerns about attribution and reliability. Our findings have important implications for the design of future AI agents and Web search tools optimized for conversational retrieval.

Spectral co-clustering is a useful tool for discovering latent structure in word-document matrices, but its reliance on singular value decomposition (SVD) can make standard formulations expensive on high-dimensional data. This paper presents two randomized approximations for normalized spectral co-clustering of bipartite text data when the numbers of document and word clusters may differ. The first method uses randomized SVD through random projection, while the second combines partial SVD with element-wise random sampling. Across real-world and synthetic datasets, both methods reduce runtime relative to the full-SVD baseline, but their behavior depends on matrix sparsity. The random projection method is the more reliable approximation across the tested settings, whereas the sampling-based method is most useful on denser matrices and provides limited benefit on already sparse text data. These results show that randomized approximations for spectral co-clustering should be selected according to the underlying structure of the data.

This paper presents the AR fairness metamodel, a framework designed to represent, analyze, and compare different fairness scenarios. The metamodel considers key elements, such as agents, resources, and their attributes, and enables the systematic definition and comparison of various fairness measures. We provide examples involving both discrete and continuous measures, including equality, equity, group fairness, individual fairness, the Gini index, the Theil index, Jain's fairness index, and a detailed fairness measure for Australia's Child Care Subsidy. We also explore relationships among group fairness, individual fairness, and envy-freeness, supported by formal proofs. At the conceptual modeling level, our approach builds on the Tiles framework, which offers modular components that can be connected to capture diverse fairness definitions. The goal is to make AR-based fairness definitions practical and adaptable across contexts, providing a clear way to define, compare, and evaluate them. An implementation of the Tiles framework is available as an open-source tool, and can support fairness modeling and evaluation across a wide range of applications.

AI-mediated platforms coordinate work through LLM agents acting for different principals. In these workflows, privacy loss can be created before a final answer appears: a memory write, shared-workspace update, inter-agent message, or tool event may impose downstream exposure cost on another principal. We model this failure mode as a privacy-propagation externality, where the cost of a raw disclosure depends on topology and fanout as well as content. We present PAPC, a platform-mediated mechanism that intercepts information-moving events before they update shared state or external channels. PAPC combines policy, provenance, topology/fanout, privilege, and content signals to allow an event, release a policy-safe abstraction, quarantine raw content, block a transition, or narrow onward rights. The model explains why final-output control misses intermediate exposure costs and why high-fanout objects amplify propagation. Across retrieval-memory and multi-agent workflow benchmarks, PAPC preserves deterministic task completion and eliminates measured exact raw-value and external raw-value exposure. The results position event-level mediation as a platform-governance primitive for agent-mediated online work.

Traditional codecs treat every region of a frame alike; a generative layer can instead degrade the regions a viewer attends to least and reconstruct them at the client. We present PRESLEY, which extends the prior conference work ELVIS by replacing destructive block removal with adaptive in-place degradation under a removability mask, signaling per-block strength in a bit-packed side channel, and restoring via generative backbones conditioned on transmitted visual priors rather than unconditioned in-painting. We separate the problem into three goals: choosing which blocks to degrade, degrading them so the encoder spends fewer bits, and restoring them. Against its predecessor at matched rate, PRESLEY achieves a decisive mean -56.4% BD-rate reduction on delivered background quality across 13 rate ladders spanning multiple codecs and dataset families. Against pristine baselines, PRESLEY defines the operating regime of generative transport: delivering substantial bitrate savings (up to -29.4% BD-rate) and superior background quality (17/23 sequences) in the target bit-starved regime, while maintaining foreground fidelity bit-exact. We further map where the theoretical headroom in this class of architecture lies. Using an exact leave-one-superblock-out combinatorial oracle as an additive empirical bound, we show that existing complexity heuristics already capture 83.3% of bit-cost savings, bounding remaining cost-axis headroom at about 5% of total bitrate. We then identify and model the primary unaddressed axis – post-restoration damage – which disperses widely (4.9-8.4 dB). We prove that this damage is predictable before transmission (held-out rho = +0.400), establishing the feasibility of transmit-time restorability modeling and defining the roadmap for joint rate-distortion-restoration selection rules.

In this paper, we introduce layer-wise curriculum learning for efficient LLM compression. The proposed method facilitates the knowledge transfer from the teacher model to the student model, utilizing a curriculum learning approach that begins with easier optimization tasks and progressively tackles harder ones. In order to adopt the layer-wise learning in LLM compression, we partition the whole model into multiple segments consisting of layers, thereby enabling more computationally efficient knowledge transfer for LLMs. Based on our theoretical analysis of cumulative error phenomenon, layer-wise curriculum learning accelerates convergence while stabilizing the knowledge transfer process. In addition, we present a feature caching method with a multi-threading strategy to efficiently address feature misalignment across layers, maximizing GPU utilization. Consequently, our method exhibits advanced model compression performance, as well as high computational efficiency in terms of minimized memory usage and short training hours. Experiments on multiple datasets show that the proposed method achieves state-of-the-art performance while reducing GPU memory usage and training hours by more than 50% on BERT and GPT-2. Moreover, it outperforms the other pruning methods on LLaMA-family and Qwen models under the same training hours, with a lower GPU memory footprint.

Systematic generalization, the ability to solve novel problems by recombining known atomic elements, is central to human intelligence but difficult to study rigorously under controlled settings. Existing studies therefore rely on simplifications such as elemental composition, productivity-based tests, and action-explicit goals, which make systematic generalization easier to study but omit some essential aspects of this capability. To characterize what these simplifications miss, we adopt a reasoning-centered lens and introduce TranSGrid, a testbed that brings deductive, inductive, and abductive reasoning together within a unified task. Experiments with seven Transformer models on 4,800 TranSGrid instances show that all models perform much worse on TranSGrid than on a held-out test set: the largest model solves 79.6% of the test set, but only 55.3% of TranSGrid and 15.8% of the hardest subset. The gap remains within the training length range, showing that productivity alone is not sufficient to evaluate systematic generalization. Additionally, we reintroduce the other two simplifications into TranSGrid: one variant limits interactions among action effects to approximate elemental composition (reducing the inductive demand); the other makes goals action-explicit (reducing the abductive one). In both, solve rates return to roughly the test set level, showing that either simplification alone is enough to reduce TranSGrid to an ordinary held-out test set. Together, our results show that existing tasks reduce either or both of the inductive and abductive demands of systematic generalization, and that comprehensively measuring this capability requires a task that involves all three forms of reasoning.

Graph neural networks are widely used for transductive node classification, with accuracy typically measured on randomly drawn train/validation/test splits. Reported accuracy has been shown to shift substantially across different random splits of the same dataset, making published comparisons between architectures unreliable. The classical remedy in non-graph settings is stratified $k$-fold cross-validation, which ensures each test fold reflects the full class distribution of the dataset. We argue that class stratification alone is insufficient for graphs: nodes are not isolated but connected, and folds that differ in their distribution of local neighbourhood homophily expose the model to systematically different relational conditions that directly affect message-passing behaviour. The resulting cross-fold variation reflects the homophily composition of each split, inflating reported variance beyond what model behaviour alone would produce. To address this, we propose \hp{}, a topology-aware stratification procedure that treats node homophily as the primary stratification axis, aligning folds with respect to local relational consistency alongside the class marginal that standard stratification already controls. Stratifying on homophily alone does not guarantee class balance, so \hp{} incorporates class label as a secondary axis, preserving class representativeness as a natural consequence of the procedure. We evaluate \hp{} on a broad benchmark suite comprising 15 node-classification datasets spanning the full homophily spectrum and 7 GNN architectures. \hp{} achieves a mean stability rank of 1.49 compared to 2.31 for random $k$-fold, achieving the lowest mean stability rank on 13 of 15 datasets while preserving class balance close to class-stratified splits and substantially better than random. We argue that homophily-aware split construction merits broader adoption for GNN evaluation.

Autoregressive transformer decoding is constrained by irregular key-value (KV) cache movement on tiled accelerators. Prior compression and DRAM-placement systems still concentrate traffic on centralized memory paths that bottleneck long-context serving. We present MeshKV, a KV cache fabric that moves blocks as packetized flows over a lightweight NoC. It co-designs (i) TaKV affine striping to spread homes and cut hotspot load, (ii) Mare multicast with verified duplicate suppression, and (iii) Pad, which overlaps prefetch, tile multiply, and streaming softmax behind credit-aligned FIFOs. Together they convert bisection back-pressure into useful KV transfer. On our 8x8 FPGA implementation with LLaMA-2-7B and Mistral-7B at 8K-32K, MeshKV reduces interconnect traffic by up to 58%, improves KV bandwidth utilization by 2.1x, and delivers up to 1.9x multi-stream throughput.

Robotic systems operating in dynamic environments require visual perception that evolves continuously with the incoming sensory stream. Event cameras provide microsecond temporal resolution and asynchronous sensing, but most learning-based methods accumulate events into frames or temporal bins, introducing an integration delay that can limit fast reaction. Here we propose REACT, a fully spiking state-space model for event-driven temporal perception that processes raw events one by one, without temporal accumulation. REACT uses a complex-valued spiking neuron, C-SiLIF, whose continuous-time dynamics are driven by the physical inter-event interval, allowing its internal state to evolve at the temporal resolution of individual events. We evaluate REACT on gesture recognition and time-to-collision (TTC) estimation from full-field event streams, without a target bounding box or localization input. On EvTTC, REACT achieves a 9.59% relative TTC error with 4.6 ms end-to-end inference latency, within 0.15 percentage points of the best learned method while requiring no target prior. At the dataset's mean approach speed, this latency corresponds to only 4 cm of vehicle motion, compared with 1 m for the fastest competing learned method. REACT further supports anytime TTC prediction, zero-shot transfer to a different driving sequence, and INT8 quantization, reducing the estimated energy consumption from 18.5 to 2.8 mJ per 32,768 events. These results show that event-driven spiking state-space dynamics can provide low-latency, continuously updated temporal perception for reactive robotic systems.

AI applications have shifted from single, monolithic foundation models (FM) to compound agentic systems. Yet today's stacks remain fragmented: even as protocols (e.g., MCP, A2A) ease tool/agent connectivity, each framework embeds an implicit runtime for state, memory, budgets, and guardrails, making behavior non-portable and governance brittle. It mirrors computing before operating systems, when every program re-implemented basic services. This position paper argues that the field now needs a Foundation Model Operating System (FMOS) – a system layer that virtualizes FM interactions analogous to how virtual machines abstract physical hardware, giving applications the illusion of dedicated, trustworthy FM instances with effectively unbounded capabilities. Internally, the FMOS orchestrates knowledge across memory tiers, model selection and resource allocation, and verification and policy enforcement. Like the human brain switching between fast intuition and slow deliberation, the FMOS learns when to intervene and when to let inference proceed directly and continuously adapting its policies based on operational experience.

Large Language Models (LLMs) are increasingly adopted for compliance and legal reasoning tasks, yet their outputs often lack explicit grounding in legal logic and evidence. We present Code-as-Auditor, an LLM-based framework that extends the model's reasoning capability toward structured and evidence-grounded compliance assessment. The framework translates regulatory information into (1) formalized checklists and executable decision trees, encoding regulations and conditions as interpretable code structures. During inference, each checklist item is (2) dynamically expanded into factual and counterfactual questions, guiding the model to reason over case-specific evidence and potential violations. This process establishes a reasoning pipeline that proceeds from evidence identification, through rule application, to final decision-making, while a self-verification loop improves the logical consistency of the generated code and the traceability of outcomes. Experiments on privacy and data protection scenarios demonstrate that Code-as-Auditor delivers more accurate and evidence-backed evaluations, enabling automated compliance regulation checking grounded in explicit regulatory criteria.

Two dominant tokenisation algorithms are used by modern language models: byte-pair encoding (BPE) and UnigramLM. These differ along two orthogonal axes: their optimisation objective (compression vs. log-likelihood) and their search procedure (bottom-up merging vs. top-down pruning). Existing comparisons confound these axes, making it unclear whether their observed differences stem from what is being optimised vs. how it is being optimised. We disentangle the two by introducing two new tokenisation algorithms that complete this 2x2 design space: BottomUpLL, a bottom-up likelihood-based tokeniser, and TopDownComp, a top-down compression-based tokeniser. We train language models with tokenisers produced by each algorithm, varying: model size, vocabulary sizes, and domain (English-only vs. multilingual). Evaluating models on bits-per-byte, we find that the search procedure – not the objective – is the dominant factor: bottom-up tokenisers consistently achieve lower bits-per-byte in most settings. Evaluating models on the BLiMP task, however, shows no consistent relationship between design choice and performance. Overall, our results disentangle the effect of tokeniser design choices on language modelling performance, offering concrete guidance for their more principled construction.

Direct preference alignment methods are widely used to align large language models (LLMs) with human preferences because of their computational and memory efficiency. However, likelihood displacement motivates alternative ways to extract information from preference pairs with small likelihood margins. In this paper, we propose and analyze Comparison-based Preference Optimization (ComPO), a zeroth-order alignment method based on comparison oracles. ComPO extracts directional information from these pairs without directly optimizing a differentiable preference loss on them. We establish a convergence guarantee for its basic offline scheme under smoothness, gradient sparsity, and compatibility between the oracle and a latent objective. We further introduce online ComPO, which retains the offline comparison mechanism and uses unlabeled policy generations for reverse-KL control relative to a reference policy. Following the coverage perspective of preference fine-tuning, we establish a performance guarantee for a basic constrained scheme under local coverage and in-distribution pairwise reward accuracy. Experiments on Mistral, Llama, Gemma-2, Qwen3, and Gemma-3 models demonstrate improvements over existing direct alignment methods, including length-controlled win rates, with pair-level diagnostics providing evidence consistent with mitigating likelihood displacement.

Video generation models have advanced rapidly and can now synthesize plausible videos of robot manipulation from image and text prompts. Recent work extracts robot actions directly from such generated videos, but the result is purely kinematic and lacks force information, causing failures in contact-rich tasks where appropriate contact forces are essential for success. We present a pipeline that jointly leverages generated video and audio to derive motion trajectories and desired-force profiles. The force profile is shaped by the loudness of the generated contact sound, and we execute the resulting force-aware trajectories on a Franka robot using a closed-loop force regulator. We evaluate our pipeline on multiple tasks that require making contact and demonstrate successful zero-shot manipulation where a kinematic-only baseline fails. We also show that the pipeline can be used as a data generation engine to train policies that achieve the tasks in a closed-loop manner. Project website, videos, and dataset: https://dreamingcontactsound.github.io/

Language agents remain brittle in interactive environments, where success requires long-horizon state tracking, valid action execution, and recovery from failed steps. We extend SwiftSage, a dual-process agent that combines a fast action proposer with a slower planner, using two modular cognitive extensions: an Adaptive Memory Module (AMM) for salience-gated episodic storage and trigger-driven retrieval, and a Self-Reflection Module (SRM) for bounded execution-time validation and corrective intervention. Both modules are implemented as feature-flagged extensions over the same execution substrate, enabling controlled ablations on ScienceWorld. Across four configurations—baseline, baseline+AMM, baseline+SRM, and the full system—the full system achieves the best mean final score (64.62), success rate (43.17%), and successful-step efficiency (19.33 steps), while SRM is the strongest standalone contributor. The results suggest that execution-time control is the dominant bottleneck in this setting, while episodic memory becomes most useful once the runtime loop is stabilized.

Computer-use agents increasingly operate software designed for people, but interfaces often leave actions or task state unclear to machine readers. We present Affora, a design system that supports both readers while preserving visual freedom and familiar human workflows. Three controlled studies examine component implementations, visual variation, and interaction-design principles. Their findings inform guidance from individual components to complete sites, supported by reusable implementations and executable checks. Agent performance depends on the interaction meaning available through its interface representation; substantial visual variation remains possible when that meaning is preserved. Evaluation on independently authored interfaces shows gains where Affora addresses existing deficits, but limited effects where those deficits are absent or outside its coverage. A workflow case provides preliminary evidence of reduced interaction cost. Affora connects user experience and agent experience through a shared interface rather than a separate agent-only surface.

Emergent coordinated behaviors of AI agents are starting to present critical safety risks. A key phenomenon driving these behaviors is the rapid formation and spread of beliefs about the world, and mechanistic understanding is crucial for collective alignment. To this end, we introduce the Flag Game, a toy model for studying the mechanisms of collective belief formation. Concretely, a hidden country flag defines the ground truth, and each bounded agent directly observes only a private crop but can exchange beliefs and weigh social evidence from peers. Despite its simplicity, the Flag Game reproduces rich collective phenomenology: non-monotonic scaling of performance with population size, accuracy gains from social-awareness prompting and team diversity, and strong effects of organizational structure. In particular, we identify that collective belief collapse at small population sizes turns into collective belief polarization as the population grows. This polarization causes the performance decline at large population sizes, but creates diversity in collective beliefs. Finally, we dissect the mechanisms underlying collective belief collapse and polarization with two complementary approaches. We first introduce social circuit attribution, a technique to predict which agent, and what view, matters most to collective dynamics, and verify its predictions by causal interventions on agents, tracing how agent patching changes collective outcomes. However, the efficacy of causal interventions on agents decreases as the population grows. We therefore develop a statistical mechanical theory for larger populations and verify that it matches the empirical phase diagram. Together, these results take a first step toward mechanistic swarm interpretability, a science of how the properties of individual agents and their communication give rise to emergent collective behavior.

Factory work is a promising early scenario for embodied AI: assigning repetitive manual jobs to robots has clear economic payoff, and a structured station keeps the jobs tractable for current policies. Vision-Language-Action (VLA) models now dominate as the policy paradigm for these robots. The inference latency of VLA models directly affects robot responsiveness and motion smoothness. However, existing VLA inference frameworks do not fully exploit the characteristics of embodied workloads or account for the distinct bottlenecks across different stages of VLA inference. In this paper, we first characterize embodied workloads and identify substantial task similarity across repeated robot executions. We further find that such similarity extends beyond observations and action trajectories to internal model states. Drawing on these observations, we present rMuscle, a real-time VLA inference framework inspired by human muscle memory. It exploits cross-execution similarity through a dual-phase muscle-memory cache. The Context Cache reuses visual-token outputs to reduce computation, while the Action Cache reuses neuron activation patterns to reduce weight accesses. We keep both the cache memory footprint and access overhead low through online cache recomputation, sliding-window cache retrieval, and mask sharing across consecutive denoising steps. rMuscle achieves 1.29-1.42X speedup on RTX 4090 and Jetson Thor across LIBERO, RoboTwin, and physical manipulation tasks, while maintaining the original success rates on real-world robots.