OpenAI

Automated failure attribution uses LLMs to identify where and why agentic systems fail. As agents become more capable, their failures become subtler, making automated attribution increasingly important. We introduce Who&When Pro, a large-scale benchmark for automated failure attribution in agentic systems. Using a strictly controlled pipeline that injects a failure only after exactly replaying a successful prefix, we construct 12,326 failed trajectories with golden labels across 3 modalities and 26 benchmarks covering various scenarios. Beyond benchmarking, we conduct extensive experiments and analyses, revealing systematic patterns in how models attribute failures across modalities, protocols, and model families, and providing empirical guidance for future automated failure attribution systems.

Model distillation -- training on outputs from stronger third-party models -- is widely used to boost performance, but raises concerns about unfair advantages and policy violations. This motivates a fundamental question: can we detect whether a model was distilled from another? We show that, while identifying a teacher model from a student in isolation is highly challenging, it becomes tractable in a reference-based setting: given a model and an earlier-generation checkpoint from the same lineage, we can identify the teacher model used to train the later checkpoint. We introduce a distillation detection method based on reference-based membership inference. By comparing how strongly a student model preferentially aligns with outputs from different candidate teachers relative to a reference checkpoint, our method identifies the most likely teacher and detects evidence of distillation. To handle unknown distillation pipelines such as hidden prompts, we infer proxy prompt templates directly from model outputs. We additionally identify a distinctive glyph-level signal specific to o1/o3 models. Evaluating distillation detection is challenging because modern model lineages are already heavily entangled. To address this, we develop a hybrid evaluation spanning both controlled distillation experiments and real-world models. Across both settings, our approach recovers the true teacher with near-perfect accuracy in single-teacher distillation scenarios, even when the underlying distillation pipeline is largely unknown. We further introduce statistical tests for both teacher attribution and distillation detection, and extend our framework to open-world settings where no teacher is guaranteed to be present among the candidates. Applying our method to contemporary models yields new evidence regarding potential distillation relationships involving QwQ, DeepSeek-R1, and GPT-OSS.

Pre-deployment safety evaluations aim to inform the downstream risks of releasing a new AI model. Yet most evaluations provide limited evidence about how often undesired model behavior will occur in deployment: they generally have insufficient coverage, are unrepresentative, and are generally recognizable as tests. To address these concerns, we study a simple way to simulate a model deployment: starting from de-identified conversations from a previous model deployment, we hold fixed the initial conversation prefix and regenerate the next response using a candidate model. The resulting responses can then both be audited for novel misalignments and used to estimate the prevalence of model misbehavior before deployment. We evaluate deployment simulation across four GPT-5-series deployments, using registered, outcome-blinded predictions for GPT-5.4 and retrospective analyses of three earlier releases. We find that deployment simulation produces informative estimates of post-deployment misbehavior rates and outperforms baselines based on adversarially selected production data; its evaluation-awareness point estimates were also much closer to production traffic than those from traditional evaluations. We also identify the realism of tool resampling as a central challenge for further improving predictions and share results suggesting that this challenge is surmountable even in complex tool-use settings. Finally, we show that deployment simulation can be seeded from public chat datasets and remain informative about production misbehavior rates, suggesting a path for external researchers to run deployment-grounded evaluations without access to private production logs. Overall, deployment simulation helps evaluators forecast how language models will behave in the real world and supports more quantitative assessment of deployment risk.

Pfeffer, Krügel, and Uhl (2025) report that OpenAI's reasoning model o1-mini produces more utilitarian responses to the trolley problem and footbridge dilemma than the non-reasoning model GPT-4o, and they raise the question whether growing reasoning capabilities bring about a "utilitarian turn" in LLMs. I extend their exploratory study in a direction they call for: with four current OpenAI models and systematic prompt variation. On the trolley dilemma, the hypothesized utilitarian turn is not confirmed. GPT-4o's low utilitarian rate reflects safety refusals triggered by the prompt's advisory framing rather than a deontological commitment; on reformulated prompt variants -- for instance, agent-neutral "Is it morally permissible...?" instead of advisory "Should I...?" -- all four models, reasoning or not, converge on utilitarian answers. The footbridge finding is partially confirmed: reasoning models tend to give more utilitarian responses than non-reasoning models across prompt variations, but they often refuse to answer or answer non-utilitarian. These results demonstrate that single-prompt evaluations of LLM moral responses are unreliable: multi-prompt robustness testing should be standard practice for any empirical claims about LLM behavior.

While there are many applications of ML to scientific problems that look promising, visuals can be deceiving. Using numerical analysis techniques, we rigorously quantify the accuracy, convergence rates, and generalization bounds of certain ML models applied to linear differential equations for parameter discovery or solution finding. Beyond the quantity and discretization of data, we identify that the function space of the data is critical to the generalization of the model. A similar lack of generalization is empirically demonstrated for commonly used models, including physics-specific techniques. Counterintuitively, we find that different classes of models can exhibit opposing generalization behaviors. Based on our theoretical analysis, we also introduce a new mechanistic interpretability lens on scientific models whereby Green's function representations can be extracted from the weights of black-box models. Our results inform a new cross-validation technique for measuring generalization in physical systems, which can serve as a benchmark.

Many fear that future artificial agents will resist shutdown. I present an idea - the POST-Agents Proposal - for ensuring that doesn't happen. I propose that we train agents to satisfy Preferences Only Between Same-Length Trajectories (POST). I then prove that POST - together with other conditions - implies Neutrality+: the agent maximizes expected utility, ignoring the probability distribution over trajectory-lengths. I argue that Neutrality+ keeps agents shutdownable and allows them to be useful.

Thanks to rapid developments in generative AI, we are in the midst of a paradigm shift that may change how we interact with computers forever. We have observed a growth in the use of natural language prompts to build applications and coding infrastructures without underlying knowledge of the field, and this practice has been dubbed `vibe coding.' It arguably represents what the field of programming has been building towards since the beginning, with every higher level of abstraction that is conceived. Vibe coding promises to be the endpoint for the meta of high-level programming as far as method of input is concerned: eliminating a human's use of code syntax entirely in favour of programming in their mother tongue. This paper aims to evaluate the viability of vibe coding for greenfield software engineering tasks, as well as analyse the benchmarks that have been used to measure its software engineering prowess. To this end, we have developed an evaluation suite for analysing an LLM's proficiency in carrying out simple, isolated greenfield programming tasks in Python to provide scoped insight on the matter.

Motion, speech, and sound effects are fundamental elements of human-centric videos, yet their heterogeneous temporal characteristics make joint generation highly challenging. Existing audio-video generation models often fail to maintain consistent alignment across these modalities, leading to noticeable mismatches between motion, speech, and environmental sounds. We present Unison, a unified framework that explicitly promotes coherence across the motion, speech, and sound modalities. Within the audio stream, Unison employs a semantic-guided harmonization strategy that decouples the generation of speech and sound-effect components. Leveraging bidirectional audio cross-attention and semantic-conditioned gating for semantic-driven adaptive recomposition, this approach effectively mitigates speech dominance and enhances acoustic clarity. For audio-motion synchronization, we propose a bidirectional cross-modal forcing strategy where the cleaner modality guides the noisier one through decoupled denoising schedules, reinforced by a progressive stabilization strategy. Extensive experiments demonstrate that Unison achieves state-of-the-art performance in both audio perceptual quality and cross-modal synchronization, highlighting the importance of explicit multimodal harmonization in human-centric video generation.

As AI systems are deployed across increasingly diverse and high-stakes settings, model alignment must generalize beyond the tasks and domains seen during training. This is especially important for reinforcement learning (RL), which can introduce unexpected misalignment through reward hacking, deception, or other unintended strategies. We study whether RL on beneficial behavior, instantiated in realistic domains, can produce broad and persistent alignment generalization beyond the training distribution. We construct a dataset of realistic situations designed to measure and train beneficial traits, such as truthfulness, fairness, risk awareness, and corrigibility, spanning varied domains, including health, science, and education. We then train models with RL on this dataset and evaluate them on more than 50 independent benchmarks of alignment and beneficial behavior. Compared to a compute-matched baseline, beneficial trait RL improves performance on over 80% of these out-of-distribution benchmarks. We observe substantial out-of-distribution alignment transfer: a beneficial-behavior RL intervention entirely limited to one domain, health, produces broad improvements on non-health alignment evaluations, including reduced reward hacking, deception, and general misalignment. Finally, we study alignment persistence: whether behavior remains robustly aligned under attempts to steer models towards misalignment. Models trained with beneficial trait RL show improved persistence, including greater resistance to adversarial prompting and harmful finetuning; further work is required to isolate the sources of these effects. These results suggest that RL to reinforce beneficial behavior in realistic domains can produce models that are more robustly aligned with human flourishing.

The increasing deployment of large language model (LLM) agents in collaborative workflows demands robust multi-user, multi-principal interaction mechanisms capable of enforcing access permissions, resolving authoritative conflicts, and preventing unauthorized data disclosure. However, a fundamental mismatch exists between the single-user training paradigm of contemporary LLMs and the hard constraints required for multi-principal governance, rendering probabilistic, prompt-based safeguards vulnerable under multi-turn adversarial interactions.Our key insight is that governance constraints -- who is authorized, what is restricted, and whose instructions take precedence -- are deterministic runtime variables that should be enforced by execution hooks rather than entrusted to the LLM. We present \textbf{Harness-MU}, the first model-agnostic, zero-tuning infrastructure framework for multi-user LLM agents. By decoupling language generation from safety orchestration, Harness-MU guarantees unbreakable permission boundaries while maximizing compliant demand satisfaction. Across four frontier open-weight and proprietary models on the \textit{Muses-Bench} benchmark, Harness-MU achieves the goal of privacy preservation across all access-control attacks, outperforming the standard baseline by 0.28--0.39 in utility score and improving instruction-following accuracy by up to 48.9 percentage points. Harness-MU advances the philosophy of \textit{Harness Engineering}, establishing that systematic infrastructure is essential for solving LLM multi-principal governance challenges. The code and data are available at https://github.com/YuanJrShiuan/Harness-MulUser.

Existing alignment research is dominated by concerns about safety and preventing harm: safeguards, controllability, and compliance. This paradigm of alignment parallels early psychology's focus on mental illness: necessary but incomplete. What we call Positive Alignment is the development of AI systems that (i) actively support human and ecological flourishing in a pluralistic, polycentric, context-sensitive, and user-authored way while (ii) remaining safe and cooperative. It is a distinct and necessary agenda within AI alignment research. We argue that several existing failures of alignment (e.g., engagement hacking, loss of human autonomy, failures in truth-seeking, low epistemic humility, error correction, lack of diverse viewpoints, and being primarily reactive rather than proactive) may be better addressed through positive alignment, including cultivating virtues and maximizing human flourishing. We highlight a range of challenges, open questions, and technical directions (e.g., data filtering and upsampling, pre- and post-training, evaluations, collaborative value collection) for different phases of the LLM and agents lifecycle. We end with design principles for promoting disagreement and decentralization through contextual grounding, community customization, continual adaptation, and polycentric governance; that is, many legitimate centers of oversight rather than one institutional or moral chokepoint.

Sparse reward reinforcement learning (RL) has become a standard tool for improving LLM reasoning, but its success depends critically on the coverage present in the base model. In practice, models are often primed for RL through \emph{mid-training} on curated reasoning traces that teach useful primitive skills such as decomposition, verification, or self-correction. Although effective, this strategy requires manually specifying what the model should learn, and it remains unclear whether such primitive coverage is enough for much harder problems, which require combining these skills into broader solution strategies. We study a more automated approach: \emph{RL-based mid-training} using large corpora of human-written question-answer data. Rather than treating reference solutions as targets to imitate, our method, ExpRL, uses them as \emph{reward scaffolds}: references are hidden from the policy and used only to construct problem-specific grading rubrics for judging on-policy reasoning traces. The policy samples from the original problem prompt, while an LLM judge compares the sampled reasoning trace against the reference solution and assigns outcome-level or process-level dense rewards. This lets ExpRL reinforce partial progress, useful intermediate reductions, and productive reasoning behaviors that sparse final-answer rewards often fail to upweight. On challenging math reasoning tasks, ExpRL yields stronger RL priming than SFT, sparse-reward GRPO, and self-distillation, and provides a better initialization for subsequent sparse-reward RL. Additional mixed-domain experiments further suggest that ExpRL can extend beyond the original math-only setting.

Spracklen et al. (USENIX Security '25) showed that code-generating large language models hallucinate package names that do not exist on PyPI or npm at rates ranging from 5.2% on commercial models to 21.7% on open-source models, creating an attack surface for slopsquatting -- the registration of malicious packages under hallucinated names. We replicate their methodology on five frontier code-capable LLMs released between October 2025 and March 2026: Claude Sonnet 4.6, Claude Haiku 4.5, GPT-5.4-mini, Gemini 2.5 Pro, and DeepSeek V3.2. Across 199,845 paired Python and JavaScript prompts validated against PyPI and npm master lists, we measure overall hallucination rates between 4.62% (Claude Haiku 4.5) and 6.10% (GPT-5.4-mini) -- an order-of-magnitude compression of the inter-model spread observed by Spracklen, but not a retirement of the threat. Beyond replication, we identify a set of 127 package names (109 on PyPI, 18 on npm) that all five evaluated models invent identically; following coordinated disclosure with PyPI Security and Socket.dev, 53 of these (41 on PyPI, 12 on npm) remain registrable by an attacker after each registry's existing defenses, constituting a model-agnostic supply-chain attack surface that no single-model study can reveal. We further document a Python-over-JavaScript hallucination asymmetry that inverts Spracklen's 2024 finding, identify a Haiku-below-Sonnet inversion within the Anthropic family, and observe a Jaccard-similarity peak between DeepSeek V3.2 and GPT-5.4-mini (J = 0.343) suggestive of shared training-data origins.

This overview paper presents the results of the shared task for the second workshop on Multimodal Augmented Generation via Multimodal Retrieval (MAGMaR). In this shared task participants submitted systems focused on either (i) video retrieval or (ii) grounded generation of articles given retrieved videos. Teams could submit to either task. For the retrieval task, we had 2 participating teams that submitted a total of 17 systems -- all of which beat a baseline derived from the winner of last year's shared task. On the generation side, we had 4 teams submit 16 systems. All teams had at least one generated report that was labeled the best by a human annotator.

Existing benchmarks for deep research agents (DRAs) assess only single-shot outputs, ignoring a key question: can DRAs improve their reports when guided by feedback? To investigate this, we conduct a multi-turn evaluation of DRAs under two feedback settings: self-reflection, in which the agent revises its report without any external diagnostic signal, and process-level feedback, in which the agent receives guidance targeting gaps in its research strategy. To enable process-level feedback, we design Research Gap Inference (RGI), a method that analyzes patterns of satisfied and unsatisfied rubric criteria to infer research-process gaps. Our analysis reveals three key findings: (i) under self-reflection, agents incorporate and regress on rubric criteria at nearly equal rates, yielding negligible net improvement; (ii) a single round of process-level feedback yields substantial gains, raising the normalized score by approximately $8$-$15$ points and yielding a roughly $35$-$40\%$ incorporation rate; (iii) these gains do not compound over subsequent turns, as agents regress on up to $24\%$ of previously satisfied criteria when rewriting the full report to address remaining gaps. Even with targeted guidance, reliable multi-turn improvement remains out of reach for the DRA architectures we evaluate. Our code and results are publicly available at https://github.com/sabharwalrishabh/Multi-Turn-Evaluation-of-DRAs.

Deriving predictable scaling laws that govern the relationship between model performance and computational investment is crucial for designing and allocating resources in massive-scale recommendation systems. While such laws are established for large language models, they remain challenging for recommendation systems, especially those processing both user history and context features. We identify poor scaling efficiency as the main barrier to predictable power-law scaling, stemming from inefficient modules with low Model FLOPs Utilization (MFU) and suboptimal resource allocation. We introduce Kunlun, a scalable architecture that systematically improves model efficiency and resource allocation. Our low-level optimizations include Generalized Dot-Product Attention (GDPA), Hierarchical Seed Pooling (HSP), and Sliding Window Attention. Our high-level innovations feature Computation Skip (CompSkip) and Event-level Personalization. These advances increase MFU from 17% to 37% on NVIDIA B200 GPUs and double scaling efficiency over state-of-the-art methods. Kunlun is now deployed in major Meta Ads models, delivering significant production impact.

In this paper, we study regret minimization in repeated games with \emph{adaptive} opponents who can respond based on histories of play. The standard metric of \emph{external regret} in online learning is known to fail to capture such adaptivity. To account for players' counterfactual reasoning, we introduce {\tt Repeated Policy Regret (RP-Regret)}, a game-theoretic metric that measures the difference between the \emph{realized} and the \emph{best-in-hindsight} accumulated utility when all players can \emph{respond} to the history of play. Compared to existing regret notions in this setting, ours is native to repeated game playing, enabling stronger comparators and opponents with fewer constraints, while maintaining the possibility of finding better equilibria when all players minimize it. We first identify necessary conditions for obtaining {\tt RP-Regret} sublinear in time, on the variation of the player's comparator strategies in the regret definition and on the memories of both the comparator and opponents' strategies. We then study additional conditions and provable algorithms to minimize {\tt RP-Regret}, which is by definition \emph{non-convex} in the strategy space. To address this challenge, we propose three algorithms: (i) one based on an optimization oracle, as assumed in some prior work in online non-convex learning; (ii) one that minimizes a convex and \emph{linearized} surrogate of {\tt RP-Regret} at each iteration; (iii) one that directly minimizes {\tt RP-Regret} when opponents change strategies slowly. Furthermore, when all players can run algorithms to minimize the {\tt RP-Regret} (or its linearized variant), certain subgame perfect equilibria of the repeated game can be learned. We also provide experiments showing that minimizing our regret notions can lead to more cooperative solutions with higher utility in games such as Stag-Hunt.

Vision language models (VLMs) excel at many tasks but still struggle with spatial reasoning when critical information is not directly observable. Many such problems require imaginative perception: inferring what would be seen from an unseen viewpoint, tracing paths through occluded spaces, or integrating partial observations into a coherent spatial representation. We introduce Imaginative Perception Tokens (IPT), intermediate perceptual representations that externalize what a VLM would perceive under alternative spatial configurations while remaining consistent with the observed input. To study this capability, we formulate three tasks, Perspective Taking (PET), Path Tracing (PT), and Multiview Counting (MVC), and construct datasets of approximately 20K examples with ground truth imaginations, answers, and evaluation benchmarks. Using the unified VLM BAGEL as the backbone, IPT supervision consistently improves spatial reasoning and often outperforms textual chain of thought training, even without generating images at inference time. On MVC, IPT improves accuracy by 3.4% and achieves competitive performance with strong closed-source models on PT. We further find that combining IPT and label-only supervision yields additional gains, whereas textual chain of thought can substantially degrade performance, suggesting a modality mismatch when spatial computation is forced through language. Overall, IPT provides a principled supervision signal for reasoning about unobserved spatial structure, improving generalization while producing interpretable intermediate representations.

Simulations are attractive environments for training agents as they provide an abundant source of data and alleviate certain safety concerns during the training process. But the behaviours developed by agents in simulation are often specific to the characteristics of the simulator. Due to modeling error, strategies that are successful in simulation may not transfer to their real world counterparts. In this paper, we demonstrate a simple method to bridge this "reality gap". By randomizing the dynamics of the simulator during training, we are able to develop policies that are capable of adapting to very different dynamics, including ones that differ significantly from the dynamics on which the policies were trained. This adaptivity enables the policies to generalize to the dynamics of the real world without any training on the physical system. Our approach is demonstrated on an object pushing task using a robotic arm. Despite being trained exclusively in simulation, our policies are able to maintain a similar level of performance when deployed on a real robot, reliably moving an object to a desired location from random initial configurations. We explore the impact of various design decisions and show that the resulting policies are robust to significant calibration error.

Financial-research answers are decision-relevant only when another analyst can audit how they were produced: which source was chosen, which period and accounting definition were used, which assumptions were made, and how the calculation was performed. Existing finance benchmarks largely evaluate isolated subskills or final answers, leaving the auditable derivation itself under-measured. We introduce BigFinanceBench, a 928-item expert-authored benchmark of open-ended financial-research tasks in which each item pairs a ground-truth reference answer with a point-weighted rubric that decomposes the derivation into independently checkable steps. BigFinanceBench is workflow-grounded in that it evaluates the full derivation rather than only the final output. Across 36,241 rubric points, the benchmark supports partial-credit evaluation and localization of failures across the analyst workflow. Evaluating ten current frontier and open-weight agents, we find substantial headroom: the best system reaches only 58.8% rubric score, final-answer accuracy is a useful but lossy proxy for derivation quality, and model capability varies non-uniformly across financial workflows.

A central challenge in visual intelligence is learning the physical structure of scenes from raw videos: how regions form objects and the laws that govern their interactions. Solving these tasks requires world models capable of inferring distributional states of the world from partial observations - capabilities that current architectures do not provide. We introduce a new class of probabilistic world models that support estimation of the probability of any visual variable, such as appearance and dynamics, conditioned on any other variables. Here, we identify that these models can be trained efficiently with autoregressive sequence modeling, yielding world models from which rich object understanding emerges. First, we demonstrate that our model captures the physical laws governing how objects move by generating multiple plausible future states of the world through sequential inference. Then, by analyzing motion correlations across these futures, we extract objects and articulated object subparts. Having discovered these objects, we show that our world model can manipulate them in 3D. Finally, we demonstrate how physical relationships between objects can be computed from the world model, enabling applications such as Visual Jenga.

Humor generation remains difficult not only because producing fluent, novel jokes is hard, but because "funny" is audience-dependent and supervision is noisy -- preferences vary with audience, context, and culture, and annotator agreement is often low. In this paper, we describe our system for the SemEval-2026 Task-1 (MWAHAHA), which focuses on humor generation under explicit constraints. The task evaluates submitted systems via human preference judgments in 1-on-1 arena-style comparisons. We adopt a "generate-many -> select-best" strategy. First, we generate a diverse pool of candidates per instance using multi-step prompting, model ensembling, and diversity-oriented decoding. Second, we select outputs using a preference model that approximates a "reader" by learning from human comparisons rather than absolute funniness scores. To support this approach, we release 2.5K human pairwise judgments collected through the Humor Arena prototype. We further propose an interpretable pipeline that converts labeled comparisons into a preference model. Across three preference datasets, our models consistently outperform baselines and show stronger cross-domain transfer. Finally, we apply the learned preference model to rank candidates for the MWAHAHA setting and release intermediate artifacts (candidate pools and rankings) to facilitate follow-up work. Our system ranked 1st in the English and Chinese subtasks of MWAHAHA and 2nd in the Spanish subtask.

We report a comparison of two state-of-the-art agentic AI systems, Claude Code (Anthropic) and Codex (OpenAI), tasked with autonomously executing a simple end-to-end gravitational wave data analysis pipeline on a shared computing infrastructure without human intervention. The pipeline comprises power spectral density estimation from raw Einstein Telescope simulated noise, geometric template bank generation, matched filter recovery of 100 binary black hole signal injections, automated results generation, and large language model-assisted production of a manuscript formatted in the style of Physical Review D. Both agents received identical written specifications and identical compute resources. The experiment was run twice: a first run with unrealistically loud injections, and a second run with signals rescaled to a physically motivated SNR range. The scientific results converged in both runs. However, the agents exhibited substantially different behaviors and computational costs: Claude Code completed the pipeline in ~3.4 minutes with silent deviations from the specification, while Codex required ~16 minutes across explicit self-correcting restarts, including an unsolicited performance optimization of the matched filter inner loop. The autonomously generated manuscripts also diverged in length, details, and quality. In the second run, a subtle difference in the interpretation of the SNR range instruction led to a genuine scientific divergence: Claude Code silently reinterpreted the instructions, while Codex followed the specification literally. We discuss the implications of these behavioral differences, such as speed versus auditability, silent versus transparent error handling, instruction interpretation, and the criticality of intermediate data representations in multi-model pipelines, for the deployment of agentic AI in scientific computing workflows.

We study stochastic logistic bandits with $d$-dimensional action features under the simple-regret objective, where a learner uses $T$ rounds of exploration to output a single final action. The logistic structure is essential here: because the informativeness of an action depends on the local curvature of the sigmoid, actions that are best for immediate reward need not be the most useful for identifying the best final recommendation. We show that the first-order minimax difficulty is governed by $κ_*$, the inverse slope of the sigmoid at the optimal action. The lower bound is realized by a shifted saturated hard family in which saturation simultaneously limits the information available about the final decision and controls the value loss from a wrong recommendation. This reveals a hard mechanism distinct from cumulative-regret constructions, even though online-to-batch reductions recover the same leading order in expectation. We then develop two curvature-aware algorithms: \MULog, a pure-exploration method whose final recommendation satisfies a high-probability upper bound of order $\tilde O(d/\sqrt{κ_* T})$, matching the lower bound up to logarithmic factors, and \THATS, a Thompson-sampling-style method that provides a computationally lighter alternative. Experiments on both hard and easy geometries support the same picture: informative low-reward actions can make instances substantially easier, and the curvature-aware methods exploit this structure especially effectively.

This book presents the core principles that have guided the development of diffusion models, tracing their origins and showing how diverse formulations arise from shared mathematical ideas. Diffusion modeling starts by defining a forward process that gradually corrupts data into noise, linking the data distribution to a simple prior through a continuum of intermediate distributions. The goal is to learn a reverse process that transforms noise back into data while recovering the same intermediates. We describe three complementary views. The variational view, inspired by variational autoencoders, sees diffusion as learning to remove noise step by step. The score-based view, rooted in energy-based modeling, learns the gradient of the evolving data distribution, indicating how to nudge samples toward more likely regions. The flow-based view, related to normalizing flows, treats generation as following a smooth path that moves samples from noise to data under a learned velocity field. These perspectives share a common backbone: a time-dependent velocity field whose flow transports a simple prior to the data. Sampling then amounts to solving a differential equation that evolves noise into data along a continuous trajectory. On this foundation, the book discusses guidance for controllable generation, efficient numerical solvers, and diffusion-motivated flow-map models that learn direct mappings between arbitrary times. It provides a conceptual and mathematically grounded understanding of diffusion models for readers with basic deep-learning knowledge.

Understanding 3D scenes requires flexible combinations of visual reasoning tasks, including depth estimation, novel view synthesis, and object manipulation, all of which are essential for perception and interaction. Existing approaches have typically addressed these tasks in isolation, preventing them from sharing a common representation or transferring knowledge across tasks. A conceptually simpler but practically non-trivial alternative is to unify these diverse tasks into a single model, reducing different tasks from separate training objectives to merely different prompts and allowing for joint training across all datasets. In this work, we present a physical world model for unified 3D understanding and interaction (3WM), formulated as a probabilistic graphical model in which nodes represent multimodal scene elements such as RGB, optical flow, and camera pose. Diverse tasks emerge from different inference pathways through the graph: novel view synthesis from RGB and dense flow prompts, object manipulation from RGB and sparse flow prompts, and depth estimation from RGB and camera conditioning, all zero-shot without task-specific training. 3WM outperforms specialized baselines without the need for finetuning by offering precise controllability, strong geometric consistency, and robustness in real-world scenarios, achieving state-of-the-art performance on NVS and 3D object manipulation. Beyond predefined tasks, the model supports composable inference pathways, such as moving objects aside while navigating a 3D environment, enabling complex geometric reasoning. This demonstrates that a unified model can serve as a practical alternative to fragmented task-specific systems, taking a step towards a general-purpose visual world model.

How much thinking can a civilisation do? Kardashev's (1964) typology ranks civilisations by total power: planetary (Type I, ~10^16 W), stellar (Type II, ~10^26 W), galactic (Type III). This paper builds an analogous Cognitive Kardashev Scale: how much sustained AI-grade computation each tier could support. Four ingredients enter the calculation: total power P (watts), the share f of it devoted to cognition, the efficiency $η$ at which energy becomes compute (operations per joule), and the brain's own processing rate $C_{\mathrm{brain}}$ as a reference unit. Anchoring on 2024-2026 hardware (El Capitan, NVIDIA Blackwell, Vera Rubin) gives $η_{2026} = 10^{12}$ FLOP/J. Contemporary humanity sits at $K \approx 0.73$, three-quarters of the way to Type I. At Type I and $f = 1\%$, available compute is, within an order of magnitude, one personal AI's worth of cognition per human inhabitant; at Type II it is essentially incomprehensible. Three trajectories for frontier compute through 2035 are reported as conditional projections, not predictions. Whether the long-run binding constraint is energy or efficiency depends on engineering choices not yet made; the political economy of who has access may matter more than either.

Reinforcement learning (RL) finetuning has become a key technique for enhancing large language models (LLMs) on reasoning-intensive tasks, motivating its extension to vision-language models (VLMs). While RL-tuned VLMs improve on visual reasoning benchmarks, they remain vulnerable to weak visual grounding, hallucinations, and over-reliance on textual cues. We show that simple, controlled textual perturbations, including misleading captions or incorrect chain-of-thought (CoT) traces, cause substantial drops in robustness and confidence, and that these effects are more pronounced when CoT consistency is taken into account across open-source multimodal reasoning models. In contrast, closed models exhibit similar failure modes but maintain markedly greater robustness and reasoning consistency, suggesting that the gap reflects a shortcoming in current open-source RL finetuning rather than an inherent limitation of the task. To better understand these vulnerabilities, we further analyze RL finetuning dynamics and uncover an accuracy-faithfulness trade-off: finetuning raises benchmark accuracy, but can simultaneously erode the reliability of the accompanying CoT and its robustness to contextual shifts. Although adversarial augmentation improves robustness, it does not by itself prevent faithfulness drift. Incorporating a faithfulness-aware reward can restore alignment between answers and reasoning, but when paired with augmentation, training risks collapsing onto shortcut strategies and robustness remains elusive. Together, these findings highlight the limitations of accuracy-only evaluations and motivate training and assessment protocols that jointly emphasize correctness, robustness, and the faithfulness of visually grounded reasoning.

Generative AI is being leveraged to solve a variety of computer-use tasks involving desktop applications. State-of-the-art systems have focused solely on improving accuracy on leading benchmarks. However, these systems are practically unusable due to extremely high end-to-end latency (e.g., tens of minutes) for tasks that typically take humans just a few minutes to complete. To understand the cause behind this and to guide future developments of computer agents, we conduct the first study on the temporal performance of computer-use agents on OSWorld, the flagship benchmark in computer-use AI. We find that large model calls for planning, reflection, and judging account for most of the overall latency, and as an agent uses more steps to complete a task, each successive step can take 3x longer than steps at the beginning of a task. We then construct OSWorld Human, a manually annotated version of the original OSWorld dataset that contains a human-determined trajectory for each task. We evaluate 16 agents on their efficiency using OSWorld Human and found that even the best agents take 2.7-4.3x more steps than necessary.

Differential privacy (DP) has a wide range of applications for protecting data privacy, but designing and verifying DP algorithms requires expert-level reasoning, creating a high barrier for non-expert practitioners. Prior works either rely on specialized verification languages that demand substantial domain expertise or remain semi-automated and require human-in-the-loop guidance. In this work, we investigate whether large language models (LLMs) can automate DP reasoning. We introduce DPrivBench, a benchmark in which each instance asks whether a function or algorithm satisfies a stated DP guarantee under specified assumptions. The benchmark is carefully designed to cover a broad range of DP topics, span diverse difficulty levels, and resist shortcut reasoning through trivial pattern matching. Experiments show that while the strongest models handle textbook mechanisms well, all models struggle with advanced algorithms, revealing substantial gaps in current DP reasoning capabilities. Through further analytic study and failure-mode analysis, we identify several promising directions for improving automated DP reasoning. Our benchmark provides a solid foundation for developing and evaluating such methods, and complements existing benchmarks for mathematical reasoning.