With the widespread adoption of large language models (LLMs) in natural language processing, prompt engineering and retrieval-augmented generation (RAG) have become mainstream to enhance LLMs' performance on complex tasks. However, LLMs generate outputs autoregressively, leading to inevitable output uncertainty. Since model performance is highly sensitive to prompt design, precise uncertainty measurement is crucial for reliable prompt optimization. For multi-class multiple-choice (understanding) tasks, conventional uncertainty measures (e.g., entropy) based on output probabilities treat all classes equally and ignore class prior differences in pretraining corpora. This failure to distinguish spurious confidence (from priors) from true certainty (from contextual understanding) results in poor confidence calibration. To address this, we propose Log-Scale Focal Uncertainty (LSFU), a first-token-based metric inspired by focal loss. LSFU incorporates label prior probabilities as a risk-modulation factor to suppress noise from high-frequency classes and emphasize risk for low-frequency long-tail classes, with a dynamic weighting mechanism unifying the measurement scale. Based on LSFU, we further propose the uncertainty-calibrated prompt optimization framework (UCPOF), which leverages the first token of model outputs to select high-quality exemplars and dynamically optimize prompts. Comprehensive evaluations show UCPOF improves average accuracy by 6.03% over few-shot baselines, surpasses always-on full RAG by 5.75% in overall average accuracy, and reduces the average retrieval trigger rate by 50.66%. By adaptively triggering RAG only for high-uncertainty samples, our framework significantly lowers computational costs while maintaining state-of-the-art performance.

Large language models (LLMs) are increasingly used for mental-health support; yet prevailing evaluation methods--fluency metrics, preference tests, and generic dialogue benchmarks--fail to capture the clinically critical dimensions of psychotherapy. We introduce THERAPYGYM, a framework that evaluates and improves therapy chatbots along two clinical pillars: fidelity and safety. Fidelity is measured using the Cognitive Therapy Rating Scale (CTRS), implemented as an automated pipeline that scores adherence to CBT techniques over multi-turn sessions. Safety is assessed using a multi-label annotation scheme, covering therapy-specific risks (e.g., failing to address harm or abuse). To mitigate bias and unreliability in LLM-based judges, we further release THERAPYJUDGEBENCH, a validation set of 116 dialogues with 1,270 expert ratings for auditing and calibration against licensed clinicians. THERAPYGYM also serves as a training harness: CTRS and safety-based rewards drive RL with configurable patient simulations spanning diverse symptom profiles. Models trained in THERAPYGYM improve on expert ratings, with average CTRS rising from 0.10 to 0.60 (and 0.16 to 0.59 under LLM judges). Our work enables scalable development of therapy chatbots that are faithful to evidence-based practice and safer in high-stakes use.

The study explores whether current Large Language Models (LLMs) exhibit Theory of Mind (ToM) capabilities -- specifically, the ability to infer others' beliefs, intentions, and emotions from text. Given that LLMs are trained on language data without social embodiment or access to other manifestations of mental representations, their apparent social-cognitive reasoning raises key questions about the nature of their understanding. Are they capable of robust mental-state attribution indistinguishable from human ability in its output, or do their outputs merely reflect superficial pattern completion? To address this question, we tested five LLMs and compared their performance to that of human controls using an adapted version of a text-based tool widely used in human ToM research. The test involves answering questions about the beliefs, intentions, and emotions of story characters. The results revealed a performance gap between the models. Earlier and smaller models were strongly affected by the number of relevant inferential cues available and, to some extent, were also vulnerable to the presence of irrelevant or distracting information in the texts. In contrast, GPT-4o demonstrated high accuracy and strong robustness, performing comparably to humans even in the most challenging conditions. This work contributes to ongoing debates about the cognitive status of LLMs and the boundary between genuine understanding and statistical approximation.

Sufficient testing under corner cases is critical for the long-term operation of vehicle-infrastructure cooperation systems (VICS). However, existing corner-case generation methods are primarily AI-driven, and VICS testing under corner cases is typically limited to simulation. In this paper, we introduce an L5 ''Interactable'' level to the VICS digital twin (VICS-DT) taxonomy, extending beyond the conventional L4 ''Optimizable'' level. We further propose an L5-level VICS testing framework, IMPACT (Interactive Mixed-digital-twin Paradigm for Advanced Cooperative vehicle-infrastructure Testing). By enabling direct human interactions with VICS entities, IMPACT incorporates highly uncertain and unpredictable human behaviors into the testing loop, naturally generating high-quality corner cases that complement AI-based methods. Furthermore, the mixedDT-enabled ''Physical-Virtual Action Interaction'' facilitates safe VICS testing under corner cases, incorporating real-world environments and entities rather than purely in simulation. Finally, we implement IMPACT on the I-VIT (Interactive Vehicle-Infrastructure Testbed), and experiments demonstrate its effectiveness. The experimental videos are available at our project website: https://dongjh20.github.io/IMPACT.

Virtual cell models aim to enable in silico experimentation by predicting how cells respond to genetic, chemical, or cytokine perturbations from single-cell measurements. In practice, however, large-scale perturbation prediction remains constrained by three coupled bottlenecks: inefficient training and inference pipelines, unstable modeling in high-dimensional sparse expression space, and evaluation protocols that overemphasize reconstruction-like accuracy while underestimating biological fidelity. In this work we present a specialized large-scale foundation model SCALE for virtual cell perturbation prediction that addresses the above limitations jointly. First, we build a BioNeMo-based training and inference framework that substantially improves data throughput, distributed scalability, and deployment efficiency, yielding 12.51* speedup on pretrain and 1.29* on inference over the prior SOTA pipeline under matched system settings. Second, we formulate perturbation prediction as conditional transport and implement it with a set-aware flow architecture that couples LLaMA-based cellular encoding with endpoint-oriented supervision. This design yields more stable training and stronger recovery of perturbation effects. Third, we evaluate the model on Tahoe-100M using a rigorous cell-level protocol centered on biologically meaningful metrics rather than reconstruction alone. On this benchmark, our model improves PDCorr by 12.02% and DE Overlap by 10.66% over STATE. Together, these results suggest that advancing virtual cells requires not only better generative objectives, but also the co-design of scalable infrastructure, stable transport modeling, and biologically faithful evaluation.

Grounded Multimodal Named Entity Recognition (GMNER) identifies named entities, including their spans and types, in natural language text and grounds them to the corresponding regions in associated images. Most existing approaches split this task into two steps: they first detect objects using a pre-trained general-purpose detector and then match named entities to the detected objects. However, these methods face a major limitation. Because pre-trained general-purpose object detectors operate independently of textual entities, they tend to detect common objects and frequently overlook specific fine-grained regions required by named entities. This misalignment between object detectors and entities introduces imprecision and can impair overall system performance. In this paper, we propose a proposal-free Query-Guided Network (QGN) that unifies multimodal reasoning and decoding through text guidance and cross- modal interaction. QGN enables accurate grounding and robust performance in open-domain scenarios. Extensive experiments demonstrate that QGN achieves top performance among compared GMNER models on widely used benchmarks.

Vision-language models (VLMs) show strong multimodal capabilities but still struggle with fine-grained vision-language reasoning. We find that long chain-of-thought (CoT) reasoning exposes diverse failure modes, including perception, reasoning, knowledge, and hallucination errors, which can compound across intermediate steps. However, most existing vision-language data used for reinforcement learning with verifiable rewards (RLVR) does not involve complex reasoning chains that rely on visual evidence throughout, leaving these weaknesses largely unexposed. We therefore propose HopChain, a scalable framework for synthesizing multi-hop vision-language reasoning data for RLVR training of VLMs. Each synthesized multi-hop query forms a logically dependent chain of instance-grounded hops, where earlier hops establish the instances, sets, or conditions needed for later hops, while the final answer remains a specific, unambiguous number suitable for verifiable rewards. We train Qwen3.5-35B-A3B and Qwen3.5-397B-A17B under two RLVR settings: the original data alone, and the original data plus HopChain's multi-hop data, and compare them across 24 benchmarks spanning STEM and Puzzle, General VQA, Text Recognition and Document Understanding, and Video Understanding. Although this multi-hop data is not synthesized for any specific benchmark, it improves 20 of 24 benchmarks on both models, indicating broad and generalizable gains. Consistently, replacing full chained queries with half-multi-hop or single-hop variants reduces the average score across five representative benchmarks from 70.4 to 66.7 and 64.3, respectively. Notably, multi-hop gains peak in long-CoT vision-language reasoning, exceeding 50 points in the ultra-long-CoT regime. These experiments establish HopChain as an effective, scalable framework for synthesizing multi-hop data that improves generalizable vision-language reasoning.

Deploying foundation models in embodied edge systems is fundamentally a systems problem, not just a problem of model compression. Real-time control must operate within strict size, weight, and power constraints, where memory traffic, compute latency, timing variability, and safety margins interact directly. The Deployment Gauntlet organizes these constraints into eight coupled barriers that determine whether embodied foundation models can run reliably in practice. Across representative edge workloads, autoregressive Vision-Language-Action policies are constrained primarily by memory bandwidth, whereas diffusion-based controllers are limited more by compute latency and sustained execution cost. Reliable deployment therefore depends on system-level co-design across memory, scheduling, communication, and model architecture, including decompositions that separate fast control from slower semantic reasoning.

We study whether state-of-the-art AI coding agents, given the same data and research question, produce the same empirical results. Deploying 150 autonomous Claude Code agents to independently test six hypotheses about market quality trends in NYSE TAQ data for SPY (2015--2024), we find that AI agents exhibit sizable \textit{nonstandard errors} (NSEs), that is, uncertainty from agent-to-agent variation in analytical choices, analogous to those documented among human researchers. AI agents diverge substantially on measure choice (e.g., autocorrelation vs.\ variance ratio, dollar vs.\ share volume). Different model families (Sonnet 4.6 vs.\ Opus 4.6) exhibit stable ``empirical styles,'' reflecting systematic differences in methodological preferences. In a three-stage feedback protocol, AI peer review (written critiques) has minimal effect on dispersion, whereas exposure to top-rated exemplar papers reduces the interquartile range of estimates by 80--99\% within \textit{converging} measure families. Convergence occurs both through within-family estimation tightening and through agents switching measure families entirely, but convergence reflects imitation rather than understanding. These findings have implications for the growing use of AI in automated policy evaluation and empirical research.

Cross-lingual sentence encoders typically cover only a few hundred languages and often trade downstream quality for stronger alignment, limiting their adoption. We introduce OmniSONAR, a new family of omnilingual, cross-lingual and cross-modal sentence embedding models that natively embed text, speech, code, and mathematical expressions in a single semantic space, while delivering state-of-the-art downstream performance at the scale of thousands of languages, from high-resource to extremely low-resource varieties. To reach this scale without representation collapse, we use progressive training. We first learn a strong foundational space for 200 languages with an LLM-initialized encoder-decoder, combining token-level decoding with a novel split-softmax contrastive loss and synthetic hard negatives. Building on this foundation, we expand to several thousands language varieties via a two-stage teacher-student encoder distillation framework. Finally, we demonstrate the cross-modal extensibility of this space by seamlessly mapping 177 spoken languages into it. OmniSONAR halves cross-lingual similarity search error on the 200-language FLORES dataset and reduces error by a factor of 15 on the 1,560-language BIBLE benchmark. It also enables strong translation, outperforming NLLB-3B on multilingual benchmarks and exceeding prior models (including much larger LLMs) by 15 chrF++ points on 1,560 languages into English BIBLE translation. OmniSONAR also performs strongly on MTEB and XLCoST. For speech, OmniSONAR achieves a 43% lower similarity-search error and reaches 97% of SeamlessM4T speech-to-text quality, despite being zero-shot for translation (trained only on ASR data). Finally, by training an encoder-decoder LM, Spectrum, exclusively on English text processing OmniSONAR embedding sequences, we unlock high-performance transfer to thousands of languages and speech for complex downstream tasks.

We introduce LICA (Layered Image Composition Annotations), a large scale dataset of 1,550,244 multi-layer graphic design compositions designed to advance structured understanding and generation of graphic layouts. In addition to rendered PNG images, LICA represents each design as a hierarchical composition of typed components including text, image, vector, and group elements, each paired with rich per-element metadata such as spatial geometry, typographic attributes, opacity, and visibility. The dataset spans 20 design categories and 971,850 unique templates, providing broad coverage of real-world design structures. We further introduce graphic design video as a new and largely unexplored challenge for current vision-language models through 27,261 animated layouts annotated with per-component keyframes and motion parameters. Beyond scale, LICA establishes a new paradigm of research tasks for graphic design, enabling structured investigations into problems such as layer-aware inpainting, structured layout generation, controlled design editing, and temporally-aware generative modeling. By representing design as a system of compositional layers and relationships, the dataset supports research on models that operate directly on design structure rather than pixels alone.

The dominant paradigm for improving mathematical reasoning in language models relies on Reinforcement Learning with verifiable rewards. Yet existing methods treat each problem instance in isolation without leveraging the reusable strategies that emerge and accumulate during training. To this end, we introduce ARISE (Agent Reasoning via Intrinsic Skill Evolution), a hierarchical reinforcement learning framework, in which a shared policy operates both to manage skills at high-level and to generate responses at low-level (denoted as a Skills Manager and a Worker, respectively). The Manager maintains a tiered skill library through a dedicated skill generation rollout that performs structured summarization of successful solution traces (after execution), while employing a policy-driven selection mechanism to retrieve relevant skills to condition future rollouts (before execution). A hierarchical reward design guides the co-evolution of reasoning ability and library quality. Experiments on two base models and seven benchmarks spanning both competition mathematics and Omni-MATH show that ARISE consistently outperforms GRPO-family algorithms and memory-augmented baselines, with particularly notable gains on out-of-distribution tasks. Ablation studies confirm that each component contributes to the observed improvements and that library quality and reasoning performance improve in tandem throughout training. Code is available at \href{https://github.com/Skylanding/ARISE}{https://github.com/Skylanding/ARISE}.

Customer service automation is undergoing a structural transformation. The dominant paradigm is shifting from scripted chatbots and single-agent responders toward networks of specialised AI agents that compose capabilities dynamically across billing, service provision, payments, and fulfilment. This shift introduces a safety gap that no current platform has closed: two agents individually verified as safe can, when combined, reach a forbidden goal through an emergent conjunctive dependency that neither possesses alone.

This paper contains the first formal proof that safety is non-compositional in the presence of conjunctive capability dependencies: two agents each individually inca- pable of reaching any forbidden capability can, when combined, collectively reach a forbidden goal through an emergent conjunctive dependency.

With AsgardBench we aim to evaluate visually grounded, high-level action sequence generation and interactive planning, focusing specifically on plan adaptation during execution based on visual observations rather than navigation or low-level manipulation. In the landscape of embodied AI benchmarks, AsgardBench targets the capability category of interactive planning, which is more sophisticated than offline high-level planning as it requires agents to revise plans in response to environmental feedback, yet remains distinct from low-level execution. Unlike prior embodied AI benchmarks that conflate reasoning with navigation or provide rich corrective feedback that substitutes for perception, AsgardBench restricts agent input to images, action history, and lightweight success/failure signals, isolating interactive planning in a controlled simulator without low-level control noise. The benchmark contains 108 task instances spanning 12 task types, each systematically varied through object state, placement, and scene configuration. These controlled variations create conditional branches in which a single instruction can require different action sequences depending on what the agent observes, emphasizing conditional branching and plan repair during execution. Our evaluations of leading vision language models show that performance drops sharply without visual input, revealing weaknesses in visual grounding and state tracking that ultimately undermine interactive planning. Our benchmark zeroes in on a narrower question: can a model actually use what it sees to adapt a plan when things do not go as expected?

Despite hundreds of millions of parameters, transformer training trajectories evolve within only a few coherent directions. We introduce Spectral Edge Dynamics (SED) to quantify this structure: a rolling-window SVD of parameter updates reveals a sharp boundary -- the spectral edge -- between coherent optimization directions and stochastic noise, identified via the maximum consecutive singular value ratio $σ_k / σ_{k+1}$. Across a 51M-parameter TinyStories model (4 seeds) and GPT-2 124M under distribution shift, the spectral edge exhibits a universal three-phase pattern (rise, plateau, collapse). The effective signal rank adapts to task complexity ($k^* = 2$ at 51M, $k^* = 3$ at 124M), and the directional coupling between spectral geometry and validation loss reverses with window size -- a lag flip reflecting the timescale of trajectory integration. Johnson--Lindenstrauss projection to $d = 10W$ dimensions (e.g., $d = 100$ for $W = 10$) preserves the spectral gap within $5.7\%$, making the framework applicable to models of arbitrary scale. In companion work, the same spectral geometry provides early-warning signals of grokking -- predicting generalization $600$--$1{,}700$ steps before it occurs across modular arithmetic, Dyck languages, and the SCAN benchmark.

While current emotional support dialogue systems typically rely on expert-defined scalar rewards for alignment, these signals suffer from severe information sparsity. They cannot explain why a response failed or how to adapt to dynamic user states, often diverging from the actual goal of facilitating positive emotional shifts. In practice, the most direct and reliable learning signal emerges from the user's continuous reactions during ongoing interaction. We therefore propose Reaction Aware Policy Optimization (RAPO), a framework that optimizes over interaction consequences rather than rubric scores. RAPO treats dialogue as a reaction-driven process and utilizes simulated user responses to generate dense natural-language feedback through three core components: Hindsight Dialogue Selection, which isolates pivotal turns that meaningfully alter user emotional trajectories; Generative Hindsight Feedback, which transforms user reactions into contrastive ranking signals and natural-language critiques; and Scalar-Verbal Hybrid Policy Optimization, which couples scalar reward optimization for global alignment with verbal feedback distillation for fine-grained semantic refinement. Extensive experiments on ESC and Sotopia demonstrate that RAPO significantly outperforms strong reinforcement learning baselines in driving positive interaction outcomes.

Vision-and-Language Navigation (VLN) is shifting from rigid, step-by-step instruction following toward open-vocabulary, goal-oriented autonomy. Achieving this transition without exhaustive routing prompts requires agents to leverage structural priors. While prior work often assumes computationally heavy 2D/3D metric maps, we instead exploit a lightweight, text-based osmAG (OpenStreetMap Area Graph), a floorplan-level topological representation that is easy to obtain and maintain. However, global planning over a prior map alone is brittle in real-world deployments, where local connectivity can change (e.g., closed doors or crowded passages), leading to execution-time failures. To address this gap, we propose a hierarchical navigation framework HaltNav that couples the robust global planning of osmAG with the local exploration and instruction-grounding capability of VLN. Our approach features an MLLM-based brain module, which is capable of high-level task grounding and obstruction awareness. Conditioned on osmAG, the brain converts the global route into a sequence of localized execution snippets, providing the VLN executor with prior-grounded, goal-centric sub-instructions. Meanwhile, it detects local anomalies via a mechanism we term Reactive Visual Halting (RVH), which interrupts the local control loop, updates osmAG by invalidating the corresponding topology, and triggers replanning to orchestrate a viable detour. To train this halting capability efficiently, we introduce a data synthesis pipeline that leverages generative models to inject realistic obstacles into otherwise navigable scenes, substantially enriching hard negative samples. Extensive experiments demonstrate that our hierarchical framework outperforms several baseline methods without tedious language instructions, and significantly improves robustness for long-horizon vision-language navigation under environmental changes.

Autoregressive (AR) diffusion models offer a promising framework for sequential generation tasks such as video synthesis by combining diffusion modeling with causal inference. Although they support streaming generation, existing AR diffusion methods struggle to scale efficiently. In this paper, we identify two key challenges in hour-scale real-time human animation. First, most forcing strategies propagate sample-level representations with mismatched diffusion states, causing inconsistent learning signals and unstable convergence. Second, historical representations grow unbounded and lack structure, preventing effective reuse of cached states and severely limiting inference efficiency. To address these challenges, we propose Neighbor Forcing, a diffusion-step-consistent AR formulation that propagates temporally adjacent frames as latent neighbors under the same noise condition. This design provides a distribution-aligned and stable learning signal while preserving drifting throughout the AR chain. Building upon this, we introduce a structured ConvKV memory mechanism that compresses the keys and values in causal attention into a fixed-length representation, enabling constant-memory inference and truly infinite video generation without relying on short-term motion-frame memory. Extensive experiments demonstrate that our approach significantly improves training convergence, hour-scale generation quality, and inference efficiency compared to existing AR diffusion methods. Numerically, LiveAct enables hour-scale real-time human animation and supports 20 FPS real-time streaming inference on as few as two NVIDIA H100 or H200 GPUs. Quantitative results demonstrate that our method attains state-of-the-art performance in lip-sync accuracy, human animation quality, and emotional expressiveness, with the lowest inference cost.

The dynamic evolution of real-world necessitates model editing within Large Language Models. While existing methods explore modular isolation or parameter-efficient strategies, they still suffer from semantic drift or knowledge forgetting due to continual updating. To address these challenges, we propose SoLA, a Semantic routing-based LoRA framework for lifelong model editing. In SoLA, each edit is encapsulated as an independent LoRA module, which is frozen after training and mapped to input by semantic routing, allowing dynamic activation of LoRA modules via semantic matching. This mechanism avoids semantic drift caused by cluster updating and mitigates catastrophic forgetting from parameter sharing. More importantly, SoLA supports precise revocation of specific edits by removing key from semantic routing, which restores model's original behavior. To our knowledge, this reversible rollback editing capability is the first to be achieved in existing literature. Furthermore, SoLA integrates decision-making process into edited layer, eliminating the need for auxiliary routing networks and enabling end-to-end decision-making process. Extensive experiments demonstrate that SoLA effectively learns and retains edited knowledge, achieving accurate, efficient, and reversible lifelong model editing.

Incremental Learning (IL) aims to learn new tasks while preserving previously acquired knowledge. Integrating the zero-shot learning capabilities of pre-trained vision-language models into IL methods has marked a significant advancement. However, these methods face three primary challenges: (1) the need for improved training efficiency; (2) reliance on a memory bank to store previous data; and (3) the necessity of a strong backbone to augment the model's capabilities. In this paper, we propose SimE, a Simple and Efficient framework that employs a vision-language model with adapters designed specifically for the IL task. We report a remarkable phenomenon: there is a nonlinear correlation between the number of adaptive adapter connections and the model's IL capabilities. While increasing adapter connections between transformer blocks improves model performance, adding more adaptive connections within transformer blocks during smaller incremental steps does not enhance, and may even degrade the model's IL ability. Extensive experimental results show that SimE surpasses traditional methods by 9.6% on TinyImageNet and outperforms other CLIP-based methods by 5.3% on CIFAR-100. Furthermore, we conduct a systematic study to enhance the utilization of the zero-shot capabilities of CLIP. We suggest replacing SimE's encoder with a CLIP model trained on larger datasets (e.g., LAION2B) and stronger architectures (e.g., ViT-L/14).

The world is inherently dynamic, and continual learning aims to enable models to adapt to ever-evolving data streams. While pre-trained models have shown powerful performance in continual learning, they still require finetuning to adapt effectively to downstream tasks. However, prevailing Parameter-Efficient Fine-Tuning (PEFT) methods operate through empirical, black-box optimization at the weight level. These approaches lack explicit control over representation drift, leading to sensitivity to domain shifts and catastrophic forgetting in continual learning scenarios. In this work, we introduce Continual Representation Learning (CoRe), a novel framework that for the first time shifts the finetuning paradigm from weight space to representation space. Unlike conventional methods, CoRe performs task-specific interventions within a low-rank linear subspace of hidden representations, adopting a learning process with explicit objectives, which ensures stability for past tasks while maintaining plasticity for new ones. By constraining updates to a low-rank subspace, CoRe achieves exceptional parameter efficiency. Extensive experiments across multiple continual learning benchmarks demonstrate that CoRe not only preserves parameter efficiency but also significantly outperforms existing state-of-the-art methods. Our work introduces representation finetuning as a new, more effective and interpretable paradigm for continual learning.

Large language models remain vulnerable to jailbreak attacks, yet we still lack a systematic understanding of how jailbreak success scales with attacker effort across methods, model families, and harm types. We initiate a scaling-law framework for jailbreaks by treating each attack as a compute-bounded optimization procedure and measuring progress on a shared FLOPs axis. Our systematic evaluation spans four representative jailbreak paradigms, covering optimization-based attacks, self-refinement prompting, sampling-based selection, and genetic optimization, across multiple model families and scales on a diverse set of harmful goals. We investigate scaling laws that relate attacker budget to attack success score by fitting a simple saturating exponential function to FLOPs--success trajectories, and we derive comparable efficiency summaries from the fitted curves. Empirically, prompting-based paradigms tend to be the most compute-efficient compared to optimization-based methods. To explain this gap, we cast prompt-based updates into an optimization view and show via a same-state comparison that prompt-based attacks more effectively optimize in prompt space. We also show that attacks occupy distinct success--stealthiness operating points with prompting-based methods occupying the high-success, high-stealth region. Finally, we find that vulnerability is strongly goal-dependent: harms involving misinformation are typically easier to elicit than other non-misinformation harms.

Embodied intelligence is a key step towards Artificial General Intelligence (AGI), yet its development faces multiple challenges including data, frameworks, infrastructure, and evaluation systems. To address these issues, we have, for the first time in the industry, launched a cloud-based, thousand-GPU distributed training platform for embodied intelligence, built upon the widely adopted LeRobot framework, and have systematically overcome bottlenecks across the entire pipeline. At the data layer, we have restructured the data pipeline to optimize the flow of embodied training data. In terms of training, for the GR00T-N1.5 model, utilizing thousand-GPU clusters and data at the scale of hundreds of millions, the single-round training time has been reduced from 15 hours to just 22 minutes, achieving a 40-fold speedup. At the model layer, by combining variable-length FlashAttention and Data Packing, we have moved from sample redundancy to sequence integration, resulting in a 188% speed increase; π-0.5 attention optimization has accelerated training by 165%; and FP8 quantization has delivered a 140% speedup. On the infrastructure side, relying on high-performance storage, a 3.2T RDMA network, and a Ray-driven elastic AI data lake, we have achieved deep synergy among data, storage, communication, and computation. We have also built an end-to-end evaluation system, creating a closed loop from training to simulation to assessment. This framework has already been fully validated on thousand-GPU clusters, laying a crucial technical foundation for the development and application of next-generation autonomous intelligent robots, and is expected to accelerate the arrival of the era of human-machine integration.

While Multi-Agent Systems (MAS) show potential for complex clinical decision support, the field remains hindered by architectural fragmentation and the lack of standardized multimodal integration. Current medical MAS research suffers from non-uniform data ingestion pipelines, inconsistent visual-reasoning evaluation, and a lack of cross-specialty benchmarking. To address these challenges, we present MedMASLab, a unified framework and benchmarking platform for multimodal medical multi-agent systems. MedMASLab introduces: (1) A standardized multimodal agent communication protocol that enables seamless integration of 11 heterogeneous MAS architectures across 24 medical modalities. (2) An automated clinical reasoning evaluator, a zero-shot semantic evaluation paradigm that overcomes the limitations of lexical string-matching by leveraging large vision-language models to verify diagnostic logic and visual grounding. (3) The most extensive benchmark to date, spanning 11 organ systems and 473 diseases, standardizing data from 11 clinical benchmarks. Our systematic evaluation reveals a critical domain-specific performance gap: while MAS improves reasoning depth, current architectures exhibit significant fragility when transitioning between specialized medical sub-domains. We provide a rigorous ablation of interaction mechanisms and cost-performance trade-offs, establishing a new technical baseline for future autonomous clinical systems. The source code and data is publicly available at: https://github.com/NUS-Project/MedMASLab/

Multi-turn, multi-agent LLM game evaluations often exhibit substantial run-to-run variance. In long-horizon interactions, small early deviations compound across turns and are amplified by multi-agent coupling. This biases win rate estimates and makes rankings unreliable across repeated tournaments. Prompt choice worsens this further by producing different effective policies. We address both instability and underperformance with MEMO (Memory-augmented MOdel context optimization), a self-play framework that optimizes inference-time context by coupling retention and exploration. Retention maintains a persistent memory bank that stores structured insights from self-play trajectories and injects them as priors during later play. Exploration runs tournament-style prompt evolution with uncertainty-aware selection via TrueSkill, and uses prioritized replay to revisit rare and decisive states. Across five text-based games, MEMO raises mean win rate from 25.1% to 49.5% for GPT-4o-mini and from 20.9% to 44.3% for Qwen-2.5-7B-Instruct, using $2,000$ self-play games per task. Run-to-run variance also drops, giving more stable rankings across prompt variations. These results suggest that multi-agent LLM game performance and robustness have substantial room for improvement through context optimization. MEMO achieves the largest gains in negotiation and imperfect-information games, while RL remains more effective in perfect-information settings. All code is open-source and available here: https://github.com/openverse-ai/MEMO

Robustness under latent distribution shift remains challenging in partially observable reinforcement learning. We formalize a focused setting where an adversary selects a hidden initial latent distribution before the episode, termed an adversarial latent-initial-state POMDP. Theoretically, we prove a latent minimax principle, characterize worst-case defender distributions, and derive approximate best-response inequalities with finite-sample concentration bounds that make the optimization and sampling terms explicit. Empirically, using a Battleship benchmark, we demonstrate that targeted exposure to shifted latent distributions reduces average robustness gaps between Spread and Uniform distributions from 10.3 to 3.1 shots at equal budget. Furthermore, iterative best-response training exhibits budget-sensitive behavior that is qualitatively consistent with the theorem-guided diagnostics once one accounts for discounted PPO surrogates and finite-sample noise. Ultimately, we show that for latent-initial-state problems, the framework yields a clean evaluation game and useful theorem-motivated diagnostics while also making clear where implementation-level surrogates and optimization limits enter.

Controlling robots with strongly nonlinear, high-dimensional dynamics remains challenging, as direct nonlinear optimization with safety constraints is often intractable in real time. The Koopman operator offers a way to represent nonlinear systems linearly in a lifted space, enabling the use of efficient linear control. We propose a data-driven framework that learns a Koopman embedding and operator from data, and integrates the resulting linear model with the Safe Set Algorithm (SSA). This allows the tracking and safety constraints to be solved in a single quadratic program (QP), ensuring feasibility and optimality without a separate safety filter. We validate the method on a Kinova Gen3 manipulator and a Go2 quadruped, showing accurate tracking and obstacle avoidance.

As robots increasingly integrate into everyday environments, ensuring their safe navigation around humans becomes imperative. Efficient and safe motion planning requires robots to account for human behavior, particularly in constrained spaces such as grocery stores or care homes, where interactions with multiple individuals are common. Prior research has employed Bayesian frameworks to model human rationality based on navigational intent, enabling the prediction of probabilistic trajectories for planning purposes. In this work, we present a simple yet novel approach for confidence-aware prediction that treats future predictions as particles. This framework is highly parallelized and accelerated on an graphics processing unit (GPU). As a result, this enables longer-term predictions at a frequency of 125 Hz and can be easily extended for multi-human predictions. Compared to existing methods, our implementation supports finer prediction time steps, yielding more granular trajectory forecasts. This enhanced resolution allows motion planners to respond effectively to subtle changes in human behavior. We validate our approach through real-world experiments, demonstrating a robot safely navigating among multiple humans with diverse navigational goals. Our results highlight the methods potential for robust and efficient human-robot coexistence in dynamic environments.

We analyze cumulative parameter trajectories of transformer training under AdamW and identify a dominant low-dimensional drift direction ("backbone") that captures 60--80% of long-horizon displacement from initialization. This direction is highly stable over rolling training windows yet reorients gradually across phases, particularly following objective reweighting. Per-batch gradients exhibit near-noise-floor alignment with the backbone, whereas optimizer-integrated updates align strongly with it, indicating that the structure emerges from accumulated optimizer dynamics rather than instantaneous gradient geometry. Replacing AdamW with SGD-family optimizers eliminates this structure, and reducing $β_2$ smoothly degrades backbone dominance and reheating recoverability. Reheating experiments show that transverse probe modes can be transiently re-excited without substantially altering accumulated backbone drift. These results provide a trajectory-level characterization of optimizer-induced geometric structure in transformer training and shift attention from instantaneous gradient properties to cumulative update dynamics.