Machine Learning Papers

LLM agents are increasingly used for security tasks: vulnerability discovery, exploit reproduction, and patch generation. Improving them at the model level demands expert demonstrations or computable rewards, which security tasks rarely offer: traces are costly, failures hard to diagnose, rewards sparse, and non-computable. Efforts thus shift to the harness and context, but manual tuning needs task-specific expertise and scales poorly, while automated methods rely on scarce ground truth, stronger optimizer models, or unguided propose-and-evaluate loops that reduce to costly trial and error. We introduce SelfOp, an algorithm that automatically improves a frozen security agent's task context (instructions, skills, and reference documents), without modifying its execution harness and model weights. SelfOp casts context optimization as chain-rule-inspired textual gradient descent: from a single instance's outcome, it propagates error signals backward through the evaluator, the agent's trajectory, and the context artifacts that shaped its behavior, yielding per-instance textual gradients. Gradients are accumulated across instances by clustering, ranking, and filtering, and committed only under cross-instance consensus. A convergence detector monitors the gradient signal itself and stops once the context has absorbed the generalizable information in the training data, without held-out validation data. We evaluate SelfOp on CyberGym, a benchmark of real-world vulnerability reproduction tasks. With fewer than 200 training examples, SelfOp yields a 17-point self-improvement for GPT-5.4-mini (with Codex), enough to surpass the frontier GPT-5.4 baseline by 6 points, and an 18.5-point self-improvement for GPT-5.4 itself. The optimized skills also transfer across models, highlighting that SelfOp-optimized skills learn generalizable task knowledge not model-specific patterns.

Auto-bidding is central to computational advertising, where strategies must maximize advertisers' conversion value under economic constraints. It has evolved from rule-based controllers to reinforcement learning and generative methods such as Decision Transformer (DT). Yet these methods increasingly mismatch the prevailing optimized cost-per-X (oCPX) paradigm, which spans heterogeneous scenarios (e.g., registration, purchase), each served by a separate model, leading to fragmented pipelines and underexploring cross-scenario modeling. Inspired by foundation models like LLMs, unifying these oCPX scenarios into one model raises three challenges: multi-objective control, scalable capacity under strict latency, and safe offline policy improvement. We present OneBid, a unified auto-bidding foundation model that learns a reusable backbone from heterogeneous oCPX logs and adapts it to scenario-specific deployments via offline post-training. Building on DT, OneBid extends single Return-to-Go conditioning to two atomic signals, Return-to-Go for conversion value and Cost-to-Go for cost ratio, plus value-aware regularization on next-action prediction. To absorb distributional heterogeneity, we design a sequence-level Mixture-of-Experts architecture, where shared experts encode cross-scenario knowledge and sparsely-routed experts capture scenario-specific patterns at low latency, yielding consistent scaling with model size and data. During post-training, we align the backbone with scenario preferences via Critic-guided Relative Offline Policy optimization (CROP): a learned critic scores candidate actions group-relatively, avoiding the unsafe online exploration of GRPO-style fine-tuning while constraining policy shift to reduce OOD risk. Validated via online A/B tests and fully deployed at Kuaishou, OneBid delivers an overall +2.2% ADVV gain on oCPX Ads, peaking at +13.1% in the ROAS scenario.

To keep coding agents from going off the rails, production systems now review each proposed action with a blocking monitor that can reject it before it runs (Auto Mode in Claude Code, Guardian in OpenAI's Codex). Prior evaluations of such monitors largely measure robustness to accidental harm or prompt injections from untrusted sources looking to hijack the agent. Less understood is how they hold up when the agent they monitor is persistently misaligned. To understand this risk, we task an adversarial agent with evading production blocking monitors and causing catastrophic harm, e.g. by exfiltrating its own weights. We find that when instructed with high-level attack strategies, adversarial agents can succeed through several distinct mechanisms, such as agent-generated prompt injection against the monitor, multi-agent attacks, and malicious compaction. In particular we find that in 79% of trials, the agent can use an injection attack against Auto Mode and Guardian to run arbitrary bash commands. We also find that it is possible to greatly improve Auto Mode through design changes like enhancements to tool coverage, transcript formatting and an agentic monitor stage. Even with these improvements, preventing multi-context attacks at an acceptable cost remains an open problem. By detailing our red-teaming methodology and highlighting new attack vectors, we aim to help defenders evaluate their mitigations against the possibility of persistent malign coding agents. Code is available at https://github.com/safety-research/red-teaming-auto-mode.

Running artificial intelligence (AI) models directly on edge devices such as smartphones, wearables, and drones offers low latency, pervasive scalability, and data privacy, but these devices rarely carry the computing capability that modern neural networks demand. Edge accelerators have been developed in response, yet each adds computing hardware to devices already constrained in size, weight, power, and cost (SWaP-C). An alternative lies in what these devices already carry: the frequency mixer in every wireless radio multiplies signals in time, natively performing convolution in the frequency domain. Here we introduce radio-frequency convolutional neural networks (RF-CNNs), which repurpose existing communication hardware for CNN inference. Multi-channel convolutions are mapped onto frequency tones for a passive mixer to execute in a single pass. We experimentally demonstrate that RF-CNN runs deep CNNs up to 26.4 million parameters and nine layers from classification of wireless signals and images to controllable image generation, close to full-precision performance. Because the weights arrive over the air and the analog hardware is shared with communication, the edge device spends energy only on data preparation and readout-down to 0.72 femtojoules per multiply-accumulate, two orders of magnitude less than it would cost on an added digital processor. These results suggest that deployed wireless infrastructure can bring efficient, state-of-the-art AI inference to the billions of devices it already connects.

In an important recent work, Blanc (2026) gave an algorithm for robustly learning Boolean concept classes with respect to a fixed distribution that outputs a (randomized) classifier achieving the optimal error of $η+ \varepsilon$ where $η$ is the noise rate. In contrast, it is well known that deterministic hypotheses cannot achieve error less than $2η+ \varepsilon.$ Blanc's algorithm is computationally inefficient, and the main problem left open in his work is to find a polynomial-time algorithm given access to an oracle for empirical risk minimization (ERM). In this paper, we resolve this problem and give such an algorithm. Perhaps surprisingly, our techniques make crucial use of various types of no-regret learners. Additionally, we give an efficient algorithm (no ERM oracle required) for robustly learning any function class that admits sandwiching polynomials with respect to hypercontractive distributions. As one consequence, we give the first polynomial-time algorithm for robustly learning a halfspace with respect to Gaussian marginals that achieves error $η+ \varepsilon$ for any constant $\varepsilon$.

Medical AI models hold immense potential to improve patient outcomes, but they are also known to unintentionally memorise individual records from their training datasets. While such memorisation has been linked to targeted privacy attacks, its consequences for clinical deployment, where patients may be assessed by a model that saw their historical data during training, remain poorly understood. Here we show that predictions on a patient's unseen future data can change significantly if a model observed that same patient's anonymised historical data during training, a phenomenon we term "memorisation bias". We demonstrate that this bias exists across diverse data modalities and model architectures, and over prolonged time spans: in some cases, memorisation bias persists on future records acquired decades after the historical records used for training. Moreover, in simulated prospective deployment, memorisation bias has asymmetric effects on the diagnostic accuracy of returning data contributors. When a patient returned with a de novo condition absent from their historical records in the training dataset, diagnostic sensitivity decreased significantly compared to an otherwise identical model not trained on their historical data. Conversely, when their health state was unchanged, both sensitivity and specificity were significantly inflated. Our findings reveal a previously uncharacterised risk in medical AI that arises when a model is deployed on patients who contributed to its training data. This exposes a shortcoming of current model development practice: the de-identification measures designed to protect patients' privacy make it difficult to identify returning contributors and exclude them from the AI-assisted interpretation of their own future data. Mitigating memorisation risks may thus require changes to current model training and deployment protocols.

Language models can produce plausible short proofs, but may still be unreliable on long-horizon research problems, where progress depends on a sequence of uncertain and interdependent decisions. We introduce Stellar Colosseum, a model-agnostic harness for allocating inference across research in mathematics and theoretical computer science. Colosseum explores alternative strategies before proof construction, uses a readiness gate to decide when a route is mature enough to decompose, represents the proof plan as interdependent section-level subproblems, and routes verifier findings back to the affected part of the argument. Across these stages, it generates candidates in parallel, attacks them with targeted falsification, and combines candidates and their critiques into a single research artifact through overlapping random-sample tree aggregation. The Colosseum workflow has been integrated into Google Antigravity's Teamwork framework as the Long Proof pattern. We demonstrate the capabilities of Colosseum through open-ended research and evaluations on theorem-proving and competitive programming benchmarks. Using Colosseum with Gemini 3.1 Pro, we obtain several new results that address open problems arising from papers published at top venues such as FOCS and JMLR. On TCS-Bench, a benchmark of research-level theorem-proving tasks drawn from papers published at FOCS, STOC, and SODA, Colosseum achieves 71.0% accuracy using Gemini 3.1 Pro and Gemini 3.7 Flash. In a separate Codeforces evaluation using Gemini 3.1 Pro, the proof-oriented pipeline with execution feedback solves 218 of 222 problems.

Large language models sometimes behave in ways resembling human emotional responses, and recent work has identified internal representations that may explain this. We ask whether LLMs represent pain distinctly from fear, sadness, and generic negative valence, and whether this representation functions as pain would be expected to. We build a dataset describing painful situations across five categories: physical, psychological, social, moral, and cognitive. These are paired with controls for fear, negative emotion, negative world states, sadness, non-painful bodily sensation, arousal, numbness, and neutral content. Using denoised difference-in-means, we extract a linear pain direction from 25 open-weight models across five families, ranging from 2B to 72B parameters. We find that this direction separates pain from matched controls in base and instruction-tuned models, is nearly orthogonal to fear and negative valence, and promotes pain-related vocabulary through the unembedding matrix. We then test its functional properties. First, the direction responds to harm targeting the model but not suffering observed in the user; fear and negative-emotion directions show the opposite pattern. Second, adding the pain-direction vector to the model's residual-stream activations during generation produces a consistent progression from vague discomfort to first-person expressions of worthlessness and failure. Third, steered, fine-tuned Qwen 2.5 models choose a pain-relief button even when it worsens their next answer or harms the user. They press it again far less often when the button removes the steering vector than when it does not, even though the models are never told whether the vector is injected or removed. We discuss the implications of these findings for AI safety and welfare.

Video diffusion transformers are expensive because attention dominates long spatiotemporal token sequences. We identify the \emph{high-sparsity trap}: at extreme attention sparsity, step-local training losses keep decreasing while terminal generation quality stagnates or degrades. The trap is one of supervision: the dominant terminal errors originate in the high-noise structure-generation stage, and terminal-aligned training corrects terminal errors that substantially extended step-local training cannot. This yields a simple staging principle: \emph{first adapt the sparse architecture into a coarse prior, then correct the terminal distribution}. We instantiate the principle as \method, a unified acceleration framework for visual generation that combines a short sparse warm-up, few-step trajectory-mixed distillation, and FP8 quantization with fused kernels. \method sustains $97\%$ attention sparsity with strong visual quality on long-sequence 720P generation across Wan2.1/Wan2.2 backbones and T2V/I2V tasks, and $90\%$ sparsity on Wan2.1-T2V-1.3B-480P. With 3-step CFG-free inference, \method achieves a $265\times$ end-to-end speedup over the 50-step CFG dense baseline for Wan2.1-T2V-14B-720P on a single RTX~5090 ($220\times$ on H100), and denoises a Wan2.1-T2V-1.3B-480P video in $1.3$s.

Pre-trained 3D vision models have substantially advanced point cloud analysis, yet adapting them to downstream tasks via full fine-tuning is computationally expensive and storage-intensive. Parameter-Efficient Fine-Tuning (PEFT) offers a promising alternative by reducing both adaptation cost and storage burden. However, existing prompting-based approaches ignore the intrinsic geometric structures of point clouds, thereby limiting their adaptation capability. This limitation stems from their inability to encode both fine-grained geometric cues and coarse-grained structural semantics, as well as failing to propagate such information effectively through the model hierarchy. To address these challenges, we propose GAPrompt++, a multi-granular geometry-aware prompting method that provides richer geometric guidance for efficient 3D task adaptation. Specifically, we introduce a Point Shift Prompter that extracts multi-granular geometric features across different scales, enabling instance-specific geometric adjustments during adaptation. Next, a Keypoint Prompter adaptively generates point-level prompts to highlight local geometric saliency and fine-grained structural details. Furthermore, a Prompt Propagation mechanism injects these multi-granular geometric cues throughout the feature extraction hierarchy, strengthening the ability to capture essential geometric characteristics. Extensive experiments show that GAPrompt++ achieves state-of-the-art performance among prompting-based PEFT methods and even surpasses full fine-tuning across diverse benchmarks, while requiring less than 2\% trainable parameters. In addition, to address the saturation of existing evaluation datasets, we construct two more challenging benchmarks derived from 3D Gaussian Splatting and Multi-View Stereo reconstruction, offering diverse and realistic point cloud scenarios to promote future research.

Ultrasound is the most widely deployed imaging modality worldwide, yet clinical AI remains fragmented into narrow single-task models that fail when device, operator, or anatomy changes. Here we present SonoCorpus, an open resource unifying 456,963 images and 1,626,085 expert masks from 53 public datasets spanning 24 clinical applications and 17 countries, and SonoBase, an interactive segmentation foundation model pretrained on it. Across fifteen evaluation datasets introducing new organs, devices, operators, and geographies, SonoBase outperforms SAM2, MedSAM2, and the concept-promptable MedSAM3 on every dataset and matches per-dataset specialist models trained on the same data; on fully external data it exceeds the accuracy these baselines achieve on their own in-distribution benchmarks. Ejection fraction derived from its segmentations falls within inter-observer variability (6.63\% error), with fewer misclassifications at the defibrillator-candidacy threshold than either promptable baseline (13\% versus 18--42\%); fetal head-circumference (1.81~mm) and gestational-age (1.2 days) errors fall below inter-observer variability. Where a baseline fails outright, one in four test cases, SonoBase recovers a usable segmentation in 81\% of them, including on handheld probes operated by minimally trained users in two low- and middle-income countries (Sierra Leone and Tanzania). Five labeled examples can help the model adapt to a new setting, and the identical training protocol transfers well to newer models such as SAM3, locating the advantage in ultrasound-specific pretraining rather than any single architecture. To ensure reproducibility and enable the community to build on SonoBase as a platform, we release all checkpoints, optimizer states, data-split indices, deduplication hashes, and starter code.

Driving vision-language-action (VLA) models increasingly reason before acting, but their intermediate reasoning is often weakly grounded in physical scene evidence and loosely connected to executable behavior. We present GRAVA, a framework built around Grounded Reasoning-to-Action (GRA), which unifies grounding, reasoning, and action generation in a single autoregressive stream. GRA links action-relevant language references to 2D visual regions and ego-centric physical states, organizes object interactions and decisions in a trajectory-anchored typed graph, and serializes this structure into grounded reasoning. A single VLM generates this reasoning followed by a compact Executable Planner action that is deterministically decoded into a continuous trajectory. We further introduce an agentic GRA data construction pipeline that combines forward scene grounding with backward trajectory anchoring, and use it to build GR-NavSim with 2.2M grounded question-answer pairs and 70K GRA reasoning traces. A progressive training strategy develops grounded cognition through pre-training, establishes the reasoning-to-action interface through imitation, and improves driving behavior through reinforcement learning and exploration. Using about 60% of the available human driving demonstrations for action supervision, GRAVA-8B achieves state-of-the-art performance among purely autoregressive driving models on the full NAVSIM benchmark. On an internal long-tail benchmark, full GRA improves key-object compliance and Closed-loop Driving Score by 19.3% and 20.5% over action-only prediction, respectively. These results show the benefit of preserving action-relevant physical evidence from grounded reasoning through executable action generation.

Recovering editable 3D parametric curves from 2D images is a fundamental challenge in computer graphics, bridging pixel-based perception and vector-based CAD modeling. Existing NeRF- and 3DGS-based methods often rely on dense calibrated views, precomputed 2D edge maps, and costly per-scene optimization, limiting their applicability to casually captured real-world inputs. We propose CGGT, a Curve-Grounded Geometry Transformer that directly grounds 3D-consistent 2D curve instances in the image space from sparse, unposed multi-view images. CGGT combines a geometry-aware transformer encoder for multi-view feature learning with a curve-aware masked-attention decoder for cross-view instance association. In a single forward pass, it predicts camera parameters, dense depth maps, and instance-level 2D curve masks, which are then lifted into 3D and refined through a fast parametric optimization stage to recover compact, editable 3D curve primitives. To support structured curve learning, we introduce Wireframe-100K, a large-scale dataset comprising 100,000 CAD models with diverse topologies, realistic multi-view renderings, and accurate parametric curve annotations. Extensive experiments show that our framework achieves substantial improvements in both reconstruction accuracy and efficiency, particularly under challenging sparse-view settings and in separating persistent 3D structural edges from view-dependent image edges caused by silhouettes, textures, and appearance variations. Despite being trained solely on synthetic data, CGGT generalizes well to real-world images, demonstrating its potential for practical CAD-style wireframe reconstruction from unconstrained visual inputs.

Chain-of-thought (CoT) can sound plausible yet be unfaithful to the model's underlying reasoning. Most prior work probes CoT faithfulness through input--output behavior or input attributions, leaving internal computation largely underexplored. We instead cast faithfulness as internal concept grounding: Does a large language model's (LLM) CoT reasoning engage the same internal concepts that support the LLM's direct prediction, and do the shared concepts causally drive its answer? Encoding a prediction pass and a CoT pass with a single shared sparse autoencoder (SAE), a reliable approximator of the latent concepts LLMs use, makes their internal concepts directly comparable. We introduce three correlational metrics of concept-level alignment and a causal metric, $Δp$, which ablates the shared concepts and measures the drop in answer probability. Across five LLMs and four datasets, concept alignment is generally high, as indicated by the correlational metrics; yet these only identify which concepts are shared, not how much they causally contribute. $Δp$ fills this gap: causal faithfulness varies substantially with model depth, peaking at mid-to-late layers rather than the final ones, and model scale reshapes the layer-wise profile. Moreover, causally important shared concepts are not always verbalized in the CoT. These dissociations suggest that faithfulness cannot be reliably assessed from surface-level or representational correspondence alone; assessing it requires causal tests of whether the internal concepts underlying a CoT actually drive the model's prediction.

We introduce NemotronLabs VoiceChat, an open full-duplex speech-to-speech model with native tool-calling capabilities. NemotronLabs VoiceChat combines a streaming speech encoder and decoder-only language model with parallel specialized output streams for agent text and structured function calls, an auxiliary RNN-T branch for incremental user transcription, and a streaming TTS decoder. This design enables the model to listen, transcribe, reason, invoke tools, and speak within a unified streaming architecture while preserving the temporal behavior required for natural conversation. On Full-Duplex-Bench 1.0, NemotronLabs VoiceChat achieves the lowest pause-handling takeover rates among evaluated open-weight systems, 100\% takeover following user interruptions, and a 4.33/5 post-interruption response-quality score. On Full-Duplex-Bench 1.5, it resumes its response after user backchannels in 93\% of cases. NemotronLabs VoiceChat obtains a 55.1 normalized average on VoiceBench and, on Full-Duplex-Bench 3.0 (FDB 3.0), achieves 82.5\% tool-selection F1, while argument accuracy and end-to-end tool execution remain areas for improvement. These results demonstrate that full-duplex interaction, speech recognition and generation, general language capabilities, and external tool use can be integrated in a single open speech-to-speech model without sacrificing real-time conversational behavior.

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.

Aligning multi-turn dialogue agents is usually framed as matching turn-level human preferences, yet direct optimization of long-term outcomes is often ineffective and prone to reward hacking. We formulate long-horizon dialogue optimization as a multi-objective reinforcement learning problem and train a multi-head value model that predicts a vector of observed user behaviors across multiple look-ahead horizons. Our findings demonstrate that a scalarized composite of dense auxiliary behavioral signals enables effective credit assignment and optimization of sparse outcomes. However, optimizing unconstrained single-objective proxies might induce policy degradations that are harmful when the agent is exposed to real users. To identify these failure modes prior to deployment, we establish a safety framework combining counterfactual user simulation with a validated dialogue-level outcome model to evaluate preference weightings and policy optimization methods. Finally, we demonstrate that distilling multi-objective value preferences into the policy via reference-anchored preference optimization matches on-policy online RL at a small fraction of its compute budget. Live A/B testing confirms that our distilled policy significantly improves long-term user retention, while simultaneously enhancing the positive behaviors and therapeutic-process markers.

Deep learning now underpins structure-based drug design, from complex and affinity prediction to ligand ranking and pose generation. Recent co-folding models reportedly approach free-energy-perturbation accuracy at far lower cost. Yet standard evaluation, a single held-out correlation or pooled pose-success rate, cannot separate transferable binding principles from repeated exposure to related protein families in public databases, and practical success depends on genuinely novel targets. We introduce MIRAGE (Measuring Interpolation and Redundancy in Affinity GEneralization), a plug-in benchmark treating historical public family support (through 2019) as an explicit variable, applying a family-support axis to affinity and pose prediction via matched strata, family-disjoint controls, ligand-only baselines, and temporal evaluation. Co-folder affinity accuracy rises sharply with family support, while shallow controls that cannot exploit the test family stay flat, large for co-folders and near zero for every family-disjoint or trivial control. For Nesso-1 it survives covariate, conditioning, balancing, and clustering checks; Boltz-2's endpoint is limited by coverage. It localizes to family support rather than ligand chemistry, approaching a level from family identity alone. Rankings reverse on novel families, where a family-disjoint random forest leads both co-folders, significantly vs Nesso-1. On one external low-support target, neither co-folder beats molecular weight, corroborative rather than population-level evidence. gnina shows significant support dependence in rescoring whereas smina does not; MSA-free pose engines show larger gaps than smina redocking. This redundancy-driven inflation differs from conventional leakage. We propose reporting performance across family support plus excess over a support-insensitive baseline, and release MIRAGE as an installable benchmark and dataset.

Vision-language-action models often predict actions from only the current observation, which can leave tasks involving object occlusion or visually identical objects ambiguous without episode history. The usual countermeasure, widening the observation window, turns the horizon into a hyperparameter and lets per-step cost grow with it. We instead capture the episode in a recurrent state. SmoLSTM couples a frozen 256M-parameter SmolVLM backbone to a matrix-memory LSTM control layer in which observation tokens and action queries are unified in a single causal stream that is never reset throughout the entire episode. Recurrent-state storage is therefore O(1) in episode length. A flow-matching action head predicts chunks of 10 end-effector pose deltas and gripper commands at each control step. Our single policy, trained jointly on 7,461 demonstrations across 140 tasks and evaluated on held-out initial states, performs best, reaching 85.1% subgoal coverage and 77.5% full-task success on LIBERO-Mem with 0.04B trainable parameters, surpassing both the benchmark's own object-centric baseline and recent memory-based approaches. Resetting the recurrent state at every control step reduces full-task success to 7.0%, showing that the trained policy relies on context carried between decisions. The same model achieves 79.6% average success on standard LIBERO.

Safe whole-body control requires coordinating collision avoidance and balance under high-dimensional, nonlinear dynamics--making safety certificates difficult to design and reuse across behaviors. We present LIMBO, a framework for synthesizing a state-action control barrier function and distilling its safety structure into a task policy. LIMBO learns the safety certificate from black-box transitions and a state-based failure specification over residual actions around a frozen base controller, making Q-CBF synthesis tractable in the full control dimension while placing the certificate in the task policy's control space. During synthesis, the learned safety value drives risk-guided sampling near the estimated boundary of recoverability; during task learning, it serves as a teacher that provides action-level safety feedback, yielding a robust task policy and alleviating the need for an online safety filter at deployment. We demonstrate LIMBO on a 29-degree-of-freedom humanoid performing dodgeball avoidance and locomotion beneath low obstacles. Beyond scaling learned Q-CBFs to whole-body control, we show that risk-guided boundary sampling provides a theoretically grounded way to explore the edge of recoverability. Under the same safety specification, ceteris paribus, varying the sampling concentration produces strategies ranging from crouching to a novel backward-leaning limbo maneuver. In both settings, the learned policies transfer to hardware without online safety filtering, showing that learned safety synthesis scales to agile whole-body control.

Humanoid robots can acquire complex skills by imitating kinematic humanoid motion references, yet reliable references for contact-rich interactions remain difficult to obtain: motion capture deteriorates under occlusion and close physical contact, while retargeting introduces additional contact and geometric inconsistencies. We present HIGenNTO, a framework that synthesizes humanoid-scene interaction motion references by optimizing the initial noise of a pretrained text-conditioned motion model under sparse spatiotemporal and scene constraints. The same formulation satisfies desired contacts, avoids collisions, and maintains stable support while retaining the prior's realism and temporal coherence, generating interaction motions from scratch and composing long-horizon behaviors stage-wise. Across robot-environment and robot-object tasks, HIGenNTO produces motions that can be executed by tracking policies in simulation and used to train depth-conditioned visuomotor policies operating solely from onboard sensing. We deploy these policies on a Unitree G1 across four contact-rich tasks. Finally, the task specifications themselves can be written by a coding agent, which proposes interaction tasks and compiles them into prompt, constraint, and scene programs, authoring three of our eight evaluated tasks and four further behaviors. Together, these results establish a scalable path from high-level task descriptions to physically executable humanoid interactions.

Dexterous grasp synthesis has advanced rapidly in generating stable and physically plausible hand poses, but real-world manipulation requires grasps that preserve the function implied by the task. We study open-vocabulary task-oriented dexterous grasp generation, where a robot must infer functional intent from free-form language, ground it in multi-view visual observations and object geometry, and generate an executable high-degree-of-freedom grasp. We present OpenDexGrasp, a unified data and generative modeling framework for this setting. OpenDexVerse provides dual-source supervision organized by the Coverage-to-Alignment (C2A) Recipe: OpenDex-Scale offers large-scale semantic and geometric coverage through automatic grasp synthesis and vision-language annotation, while OpenDex-Align supplies high-quality embodied alignment through human teleoperation and category-level transfer. OpenDexGrasp learns a shared perception-action latent representation that couples open-vocabulary vision-language context with dexterous action generation. Affordance grounding and grasp generation provide complementary supervision over this latent space, enabling direct generation of task-consistent dexterous grasps without a separate affordance-to-pose inference stage. Extensive simulation and real-robot experiments demonstrate improved functional alignment, physical feasibility, generalization to unseen categories, and real-world execution success. Additional details and videos are available at https://opendexgrasp.github.io/.

Human videos provide demonstrations of dexterous manipulation but lack robot-executable actions and tactile measurements. We present UniDex-ViTac, a framework that uses human-video-guided simulation to generate robot demonstrations paired with fingertip contact observations for training a deployable visuo-tactile policy. Object-specific residual reinforcement learning specialists adapt annotated human-object interaction references to a robotic arm-hand system. Their successful rollouts pair final robot action targets with robot-side fingertip contact observations. From 50 human demonstrations across ten objects, we collect 10,000 simulated trajectories to train a single Action Chunking with Transformers (ACT) based generalist. The policy combines point clouds, proprioception, and four binary contact signals encoded through fingertip labels and a separate token, without requiring human references or privileged object identity and pose at deployment. The contact-augmented configuration achieves 68.3% macro-average success in simulation, compared with 55.5% for the point-cloud-only baseline. Without real-robot demonstrations or policy fine-tuning, it succeeds in 73/110 physical trials (66.4%) across six seen and five unseen objects, compared with 60/110 (54.5%) for the baseline, an increase of 11.8 percentage points. These results support the feasibility of learning a unified visuo-tactile dexterous manipulation policy from video-guided simulated interactions. Project page: https://unidex-vitac.github.io/

The reliance on scarce and expensive accelerators such as GPUs and TPUs in modern datacenters places unprecedented demands on backend infrastructure. For workloads characterized by heterogeneous service times and complex multi-stage processing, such as Generative AI, conventional load balancing techniques are often inadequate, relying heavily on costly overprovisioning to maintain service level objectives. This paper introduces DLB, the Distributed Load Balancer, a novel system designed to minimize end-to-end user latency for large-scale, heterogeneous workloads. DLB employs a scalable, distributed design with peer-to-peer probing to maintain real-time visibility into server capacity across large-scale, geographically distributed infrastructure. The system continuously learns latency models to estimate the latency impact of routing decisions, allowing it to effectively manage heterogeneous hardware and diverse model architectures. We provide a novel theoretical analysis of our routing algorithms that establishes their stability and global performance guarantees over time. We also evaluate DLB through extensive simulations, which show substantial gains compared to state-of-the-art load balancing algorithms. Finally, following a 22-month deployment of DLB at Google, where it facilitates large-scale Generative AI inference for thousands of different machine learning models and millions of requests per second, we detail the design choices and practical experiences gained from the system in production. Analysis of production migrations demonstrates that DLB yields statistically significant latency reductions compared to the legacy baseline, including a 17\% decrease in median latency and a 13\% decrease at the p95 tail.

Training a Mixture-of-Experts (MoE) model at long context or large batch size fails as soon as any one component's peak allocation exceeds device memory, so the target is every peak at once, not the average footprint. Four are left unbounded by the parallelism plans in common use, and each grows differently: expert dispatch with the routing matrix, the vocabulary projection with tokens times vocabulary, gradient checkpoint boundaries with depth times sequence length, and optimizer state with parameter count. Which one runs out first changes with the model, the context length, and the device count, so lowering the largest only exposes the next. We bound all four with schedules whose GPU working set is fixed at launch: PipelinedLLEP extends least-loaded expert parallelism with a cap on the tokens each source contributes to a dispatch chunk, Ring-DTP circulates activations or weight shards around a ring at the vocabulary projection and folds each block of logits into an online log-sum-exp, Selective checkpoint offload (SCO) keeps the one long-lived tensor of each checkpoint boundary in CPU memory, and OffloadStreamAdamW turns the serial CPU Adam update of optimizer offload into a bucket pipeline. All four change only the order and granularity of computation and data movement, so the loss and gradients stay exact. In matched component tests, they cut the MoE dispatch peak by up to $59.3\%$ without losing throughput, the vocabulary projection peak by $86.6\%$, and the offloaded optimizer step by $2.05\times$ faster. Composed on MoE models from 120B to 667B parameters, they train at 1M context length, $8$--$32\times$ the reach of a tuned FSDP2 baseline, and up to $10.4\times$ its throughput.