Princeton University（普林斯顿大学）

As more capable AI models are increasingly integrated into critical computer systems, the lack of control over AI intent motivates safety mechanisms. Existing software safeguards impose only behavioral constraints that can potentially be bypassed by sufficiently intelligent models. While hardware-level safety enforcement has been recognized as an essential last line of defense, few mechanisms have been proposed beyond policy regulations on unauthorized accesses or coarse full-chip shutdown. What is missing is a fine-grained, dynamic intervention mechanism at the architecture level. In this paper, we introduce a set of microarchitecture knobs which dynamically control the available hardware resources to limit AI performance at runtime. We evaluate candidate knobs spanning the GPU memory subsystem, across capacity, bandwidth, latency and frequency dimensions, and narrow down to four strong candidates: L2 size, L2 latency, L2 bandwidth, and shared memory port access rate. To minimize new logic and extra design cost, we build all four mechanisms from well-established microarchitectural primitives: cache way masking, credit-based rate limiting, latency insertion, and bank arbitration. We show that these knobs achieve high performance sensitivity (up to 80% performance cut at 1/8 resource availability), negligible implementation cost (<~10K flip flops), fast stabilization after dynamic throttling (5-80K cycles), and minimal collateral impact on the rest of the chip. Further, multi-knob analysis reveals combinations of knobs that amplify the performance degradation beyond the effect of each knob individually, which enables a broader range of performance targets.

Vision--language models have become increasingly capable of interpreting visual scenes, but it remains unclear whether they can use information to make strategically effective decisions. We investigate this question in soccer, where models observe the seconds preceding an on-ball decision and must choose whether to shoot or pass to a specific teammate. Unlike conventional visual-understanding tasks, soccer enables decisions to be evaluated quantitatively by estimating the value of every available action. We introduce SportD, a benchmark comprising 478 on-ball decisions from the 2022 FIFA World Cup. Each model choice is evaluated against a possession-value model that estimates the action that most increases the attacking team's probability of scoring, allowing us to measure both optimal-action accuracy and the value forfeited by suboptimal decisions. Across three frontier VLMs, the best selects the highest-valued action on 31.4% of events, compared with 38.9% for the professional players, and all models incur significantly greater regret. Further analysis reveals a systematic preference for lower-variance and lower-reward actions: VLMs shoot less often and select substantially less progressive passes than either the optimal policy or the real players. The models also reproduce the player's specific action above chance even when that action is suboptimal, suggesting partial imitation of familiar play patterns rather than consistent evaluation of counterfactual alternatives. SportD provides a value-grounded testbed for measuring physical strategic reasoning in VLMs.

Data-driven acceleration of scientific computing workflows has been a high-profile aim of machine learning (ML) for science, with numerical simulation of transient partial differential equations (PDEs) being one of the main applications. The focus thus far has been on methods that require classical simulations to train, which when combined with the data-hungriness and optimization challenges of neural networks has caused difficulties in demonstrating a convincing advantage against strong classical baselines. We consider an alternative paradigm in which the learner uses a classical solver's own data to accelerate it, enabling a one-shot speedup of the simulation. Concretely, since transient PDEs often require solving a sequence of related linear systems, the feedback from repeated calls to a linear solver such as preconditioned conjugate gradient (PCG) can be used by a bandit algorithm to online-learn an adaptive sequence of solver configurations (e.g. preconditioners). The method we develop, PCGBandit, is implemented directly on top of the popular open-source software OpenFOAM, which we use to show its effectiveness on a set of fluid and magnetohydrodynamics (MHD) problems.

Automated essay scoring (AES) enables scalable assessment and timely feedback but remains challenged by transformer input-length limitations, which can cause information loss when processing long essays. This study proposes a generative AI-assisted summarization framework to improve long-form essay representation while maintaining scoring reliability. Using the ASAP 2.0 dataset, we generate controlled-length summaries with three GPT-5 variants (GPT-5, GPT-5 mini, and GPT-5 nano) and use them as inputs for downstream AES models. To preserve original writing signals, handcrafted linguistic features extracted from full essays are integrated with summary representations to form a hybrid framework. The approach is evaluated in terms of scoring performance, summarization quality, and computational cost. Scoring reliability is measured using quadratic weighted kappa (QWK), while summary quality is assessed through lexical overlap, semantic similarity, information retention, and redundancy metrics. Results show that GPT-5 mini achieves the highest agreement with human ratings, whereas GPT-5 produces the strongest summarization quality. Summary quality decreases for higher-scoring essays, indicating that more complex writing is more difficult to compress without information loss. These findings reveal trade-offs among model capacity, summary fidelity, cost efficiency, and preservation of educational constructs. This study provides an initial controlled evaluation of GPT-based summarization for AES and identifies important baselines and ablation studies required for future generalization. Overall, generative AI summarization offers a promising approach for scalable writing assessment while requiring careful validation of information preservation and fairness.

Transformer reasoning is limited by autoregressive decoding, which repeat edly compresses rich hidden computation through token space and makes it difficult for intermediate reasoning states to persist across time. We in troduce Transformers with Temporal Middle-Layer Recurrence (T2MLR), a transformers-based latent reasoning architecture that fuses a cached middle layer representation from the previous token directly into an earlier layer of the current token position, enabling abstract intermediate computation to persist across decoding steps with little inference overhead. Across natural-language pretraining and multi-hop reasoning finetuning, T2MLR consistently outperforms data- and parameter-matched Transformer base lines. Moreover, applying recurrence to only a localized middle-layer block (as little as 20% of the network) often outperforms full-layer recurrence. Im portantly, T2MLR does not require pretraining from scratch: retrofitting the recurrent pathway into an existing pretrained 1.7B Transformer and briefly finetuning substantially improves math reasoning, lowering the barrier to practical adoption. These results suggest that effective latent reasoning in Transformers does not require looping over all layers as in previous works, but can instead emerge more strongly from targeted middle-layer recurrence.

Noise injection is a well-known technique in stochastic optimization. We report its surprising effectiveness with an interleaved (on-off-on-off...) rather than the usual monotonic decay schedule. We present a theoretical analysis of noise injection, which confirms that corruption by impulse noise approximates a Jacobian regularization, whereas Gaussian noise acts as a curvature penalty. This regularization behavior has been invoked to explain why noise injection increases model robustness. But the interleaved nature of our proposed schedule produces superior results even for the optimization objective: mixing phases of noisy data permits the optimizer to escape local minima and increase exploration without the risk of catastrophically forgetting the important features from the clean data. To stabilize this training scheme against the rapid changes of the loss when switching between clean and noisy data, we introduce a gradient-norm stabilization technique that scales noisy updates based on clean gradient magnitudes. We compare this method with other common augmentation methods and find substantial improvements in corruption tolerance and robustness to real-world distribution shifts on CIFAR-100-C, ImageNet-C, and ImageNet-R for ResNet and ViT architectures, with the best results being achieved by stacking our method on top of other augmentations. Through saliency and attention maps we show that the effect of interleaved noise injection stems from penalizing the failure modes encouraged by the inductive bias of the models: impulse noise works against the locality bias of convolutional (ResNet) architectures, and Gaussian noise reduces the tendency of attention-based models to pick up large-scale spurious features. Interleaved noise injection is therefore an effective tool to improve the test performance on clean, noisy, and out-of-distribution data at essentially zero computational cost.

Reaction mechanisms consist of the step-by-step sequences of elementary reactions that explain chemical transformations. Learning the mechanism logic is therefore essential for enhancing the fundamental chemical intelligence of large language models (LLMs). The stepwise deduction of reaction mechanism aligns naturally with the reasoning paradigms of reasoning LLMs. However, current chemical LLMs primarily emphasize coarse-grained name reactions for product prediction and retrosynthesis, often leading to physical inconsistencies and hallucinations. In contrast, specialized small-scale generative models for mechanism inference typically suffer from restricted generalization capacity across diverse chemical spaces. To overcome these limitations, we built a novel, large-scale reasoning dataset of reaction mechanisms. Furthermore, we established the FukuyamaBench, a difficult benchmark derived from Fukuyama's Advanced Organic Reaction Mechanism book, to rigorously evaluate model performance on hierarchical mechanism reasoning. Our fine-tuned Qwen3-30B-A3B achieves 8.3% exact pathway match on FukuyamaBench Set~A, surpassing the specialized FlowER model (5.1%), demonstrating that mechanism-aware training substantially enhances chemical reasoning in language models.

Looped Transformers scale sequential computation by applying a compact stack of physical blocks for multiple rounds, increasing unrolled depth without increasing stored parameters. This reuse changes the residual-scaling problem: in an untied Transformer, each residual branch receives and applies its own parameter update, whereas in a looped Transformer one shared update aggregates gradients from repeated visits and is read back by those same visits in the next linearized forward pass. We formalize this tied-depth effect through a first-order perturbation bound controlled by a visit-alignment coefficient $κ_R$. The bound recovers the DeepNorm exponent when visits decorrelate, but in the conservative aligned regime it requires the exponent to increase from $1/4$ to $1/2$ as loop count grows at fixed physical depth. The resulting method, \textbf{DeepLoop}, keeps the Post-LN DeepNorm architecture and sets $α=(2N)^{1/2}$ and $β=(8N)^{-1/2}$ for unrolled depth $N$. On GPT-style looped language models at GPT-2 small and GPT-2 medium scale, DeepLoop is neutral when no physical block is revisited and improves validation loss and downstream accuracy once recurrent depth is activated. These results show that stable recurrent depth requires residual scaling rules that account for parameter visits, not only nominal layer count.

Reinforcement learning (RL) promises to expand the capabilities of language models, but it is unclear if current RL techniques promote the discovery of novel behaviors, or simply sharpen those already present in the base model. In this paper, we investigate the value of deliberate exploration -- explicitly incentivizing the model to discover novel and diverse behaviors -- and aim to understand how the knowledge in pre-trained models can guide this search. Our main finding is that exploration with a simple, principled, representation-based bonus derived from the pre-trained language model's hidden states significantly improves diversity and pass@k rates -- both for post-training, and in a novel inference-time scaling setting we introduce. For inference-time, exploration with representation-based diversity improves efficiency, consistently improving pass@k rates across a variety of models and reasoning tasks. For example, for Qwen-2.5-14b-Instruct we obtain over 50% improvement in verifier efficiency on almost all tasks. For post-training, we show that integrating this exploration strategy into an RL pipeline improves reasoning performance over that of the initial model and over standard RL post-training. For example, on AIME 2024, our post-trained Qwen-2.5-7b-Instruct's pass@80 matches the pass@256 of GRPO on the same model, demonstrating a 3x improvement in test-time sample efficiency. Overall, our findings suggest that deliberate exploration -- with the right notion of diversity -- is a practical path toward discovery of new behaviors beyond sharpening.

Reliable confidence -- the probability that a model's own answer is correct -- is essential for the trustworthy deployment of language models. Existing work has largely evaluated confidence by how well it predicts correctness and whether it is calibrated, leaving open a more fundamental question: what does the confidence signal itself represent? Answer logits may reflect a latent decision variable sufficient to compute normative confidence, or instead a heuristic preference signal that combines the available evidence in a non-Bayesian manner. We address this using statistical decision confidence (SDC), a normative framework from computational neuroscience. Treating the answer-logit difference (LD) as a candidate readout of the latent decision variable, we test the qualitative signatures predicted by SDC. Across three perceptual discrimination tasks and a memory-based decision task, spanning three multimodal non-reasoning models and one reasoning model, LD satisfied these signatures -- including the diagnostic correct/error folded-X pattern -- showing that, in these settings, answer logits behave as monotonic readouts of a latent decision variable rather than heuristic preference scores. In complex visual reasoning, LD continued to predict correctness beyond objective task difficulty, but the full geometric signatures of SDC were absent, illustrating the current boundary of the framework when explicit normative process models are unavailable. These results provide a computational account of confidence in multimodal language models, delineate when answer logits behave as readouts of a latent decision variable, and establish SDC as a unifying framework for studying confidence across biological and artificial intelligence.

We exhibit an exact correspondence between sampling with score-based diffusion models and adiabatic transport of ground states for a family of Schrödinger operators we call Score Hamiltonians, built from the learned score's quantum potential. We obtain novel density reconstruction bounds and principled annealing schedules via adiabatic theorems for Fokker-Planck equations with time-varying potentials. We find the fundamental limit of sampling is set by the ratio of squared score-matching error to Score Hamiltonian spectral gap - the inverse Poincaré constant of the data density.

This paper describes the first ChineseBabyLM challenge, which will be held in the 2026 NLPCC conference. The challenge calls for researchers to train language models from scratch with 100 million Chinese tokens and evaluates the models on 3 tracks of tasks: NLU, cognitive alignment and Hanzi knowledge. There is no restriction on tokenizer, model architecture and the number of training epochs. Details of the challenge can be found in https://chinese-babylm.github.io/.

Partial differential equations (PDEs) govern complex systems, yet neural operators often struggle to efficiently capture the long-range, nonlocal interactions inherent in their solution maps. We introduce Spectral Filtering Operator (SFO), a neural operator that parameterizes integral kernels using the Universal Spectral Basis (USB), a fixed, global orthonormal basis derived from the eigenmodes of the Hilbert matrix in spectral filtering theory. Motivated by our theoretical finding that the discrete Green's functions of shift-invariant PDE discretizations exhibit spatial Linear Dynamical System (LDS) structure, we prove that these kernels admit compact approximations in the USB. By learning only the spectral coefficients of rapidly decaying eigenvalues, SFO achieves a highly efficient representation. Across six benchmarks, including reaction-diffusion, fluid dynamics, and 3D electromagnetics, SFO achieves state-of-the-art accuracy, reducing error by up to 40% relative to strong baselines while using substantially fewer parameters.

Games have long been a microcosm for studying planning and reasoning in both natural and artificial intelligence (AI), often focusing on expert-level or even super-human play. But real life also pushes human intelligence along a different frontier, requiring people to flexibly navigate decision-making problems that they have never thought about before. Here, we use novice gameplay to study how people reason about new problem settings. Through a series of large-scale behavioral studies with over 1000 participants and 121 two-player strategic board games (almost all novel to our participants), we show that people are systematic and adaptively rational in how they play a game for the first time, or evaluate a game (e.g., how fair or how fun it is likely to be) before they have played it even once. We explain these capacities via a computational cognitive model that we call the 'Intuitive Gamer', a model based on mechanisms of fast and flat (depth-limited) goal-directed probabilistic simulation. Our work offers new insights into how people rapidly evaluate, act, and make suggestions when encountering novel problems, and could inform the design of more flexible and human-like AI systems that can determine not just how to solve new tasks, but whether a task is worth thinking about at all.

Object-Goal Navigation (ObjectNav) is a key capability for deploying mobile robots in everyday environments such as homes, schools, and workplaces. In this task, an agent must locate an instance of a target object category in previously unseen environments using only onboard perception, requiring the integration of semantic understanding, spatial reasoning, and long-horizon planning. Reinforcement learning (RL) has become a dominant paradigm for ObjectNav, yet modern systems involve numerous design choices across perception modules, policy architectures, and inference-time strategies. The relative impact of these components, however, remains poorly understood. In this work, we present a large-scale empirical study of modular RL-based ObjectNav systems. We decompose the navigation pipeline into three key components: perception, policy, and test-time enhancement, and conduct extensive controlled experiments to analyze their individual contributions. Our results suggest that improvements in perception quality and test-time strategies often yield larger performance gains than policy improvements alone, highlighting the importance of understanding how different components interact within modular navigation systems. Motivated by these findings, we introduce a unified framework for systematically studying modular ObjectNav systems. Guided by our analysis, we build an enhanced system that achieves state-of-the-art performance on the Gibson benchmark, improving SPL by 6.6% and success rate by 2.7% over prior methods. We also introduce a human expert baseline, achieving 98% success, highlighting the significant gap between current RL agents and human-level navigation. Finally, we provide practical insights and design recommendations for each module to help guide future research. Project page: https://honwang0054.github.io/What-matters-in-RL-ObjNav-web/.

Natural language data follows a power-law distribution, with most knowledge and skills appearing at very low frequency. While a common intuition suggests that reweighting or curating data towards a uniform distribution may help models better learn these long-tail skills, we find a counterintuitive result: across a wide range of compositional reasoning tasks, such as state tracking and multi-step arithmetic, training under power-law distributions consistently outperforms training under uniform distributions. To understand this advantage, we introduce a minimalist skill-composition task and show that learning under a power-law distribution provably requires significantly less training data. Our theoretical analysis reveals that power law sampling induces a beneficial asymmetry that improves the pathological loss landscape, which enables models to first acquire high-frequency skill compositions with low data complexity, which in turn serves as a stepping stone to efficiently learn rare long-tailed skills. Our results offer an alternative perspective on what constitutes an effective data distribution for training models.

As unsupervised pretraining becomes increasingly ubiquitous in reinforcement learning, a more thorough theoretical understanding of these methods becomes of equal importance to their empirical success. We focus on the setting of unsupervised learning via interaction, where the forward-backward (FB) representation learning serves as a prototypical and popular example. In this paper, we shed light on FB by formally contextualizing the method within a broader class of recent methods that use regression to obtain a low-rank approximation of a successor measure ratio. Our analysis clarifies when FB representations can exist and how the low-rank approximation converges in practice. Building upon the theory, we propose a variant of FB that is both more amenable to theoretical understanding and simpler to optimize in practice. Experiments in didactic settings, as well as in $10$ state-based and image-based continuous control domains, demonstrate that our method converges to desired representations with $10^5 \times$ smaller errors than FB, achieving $+24\%$ improved zero-shot performance on average. We also demonstrate that zero-shot policies inferred by our algorithm provide an efficient initialization if the user prefers further fine-tuning on downstream tasks. Our project website is available at https://chongyi-zheng.github.io/onestep-fb.

Compositional visual reasoning has emerged as a key research frontier in multimodal AI, aiming to endow machines with the human-like ability to decompose visual scenes, ground intermediate concepts, and perform multi-step logical inference. While early surveys focus on monolithic vision-language models or general multimodal reasoning, a dedicated synthesis of the rapidly expanding compositional visual reasoning literature is still missing. We fill this gap with a comprehensive survey spanning 2023 to 2025 that systematically reviews 260+ papers from top venues (CVPR, ICCV, NeurIPS, ICML, ACL, etc.). We first formalize core definitions and describe why compositional approaches offer advantages in cognitive alignment, semantic fidelity, robustness, interpretability, and data efficiency. Next, we trace a five-stage paradigm shift: from prompt-enhanced language-centric pipelines, through tool-enhanced LLMs and tool-enhanced VLMs, to recently minted chain-of-thought reasoning and unified agentic VLMs, highlighting their architectural designs, strengths, and limitations. We then catalog 60+ benchmarks and corresponding metrics that probe compositional visual reasoning along dimensions such as grounding accuracy, chain-of-thought faithfulness, and high-resolution perception. Drawing on these analyses, we distill key insights, identify open challenges (e.g., limitations of LLM-based reasoning, hallucination, a bias toward deductive reasoning, scalable supervision, tool integration, and benchmark limitations), and outline future directions, including world-model integration, human-AI collaborative reasoning, and richer evaluation protocols. By offering a unified taxonomy, historical roadmap, and critical outlook, this survey aims to serve as a foundational reference and inspire the next generation of compositional visual reasoning research.

When analyzing a manifold learning algorithm for data lying on a smooth, compact, connected Riemannian submanifold $(\mathcal{M}, g)$ of $\mathbb{R}^d$, a key estimate for the geodesic distance $d_g$ is that there exists $K > 0$ such that $0 \leq d_g(p, q)^2 - \|p-q\|^2 \leq K d_g(p, q)^4$ for all $p, q \in \mathcal{M}$. We observe that more generally, when $\mathcal{M}$ is equipped with a smooth symmetric divergence $D$ satisfying a non-degeneracy condition and $g$ is given by $g_p := \frac{1}{2}\mathrm{Hess}_p(D(p, \cdot))$ for all $p \in \mathcal{M}$, there exists $K > 0$ such that $\left| D(p, q) - d_g(p, q)^2 \right| \leq K d_g(p, q)^4$ for all $p, q \in \mathcal{M}$. We demonstrate that this is sufficient for the pointwise convergence of graph Laplacians constructed with $D$ and discuss examples where $D$ is given by the Sinkhorn divergence on a family of probability measures parametrized by a manifold.

We introduce EgoPolice, a carefully curated dataset of real, egocentric police-civilian interactions, sourced from publicly available body-worn camera videos. We select police-civilian action labels that are critical for police behavioral research and annotate them at a second-by-second granularity. The videos feature rapid and irregular camera motion, dense human interactions, and rare high-stakes events, making the dataset a challenging benchmark for motion-robust and context-aware egocentric perception. We provide two different tasks, classification and multiple-choice question-answering, and benchmark both open-source and closed-source models. We find that even the best video models like Gemini 2.5 Pro still struggle to accurately predict high-risk actions such as "Weapon Out". Beyond serving as a benchmark, EgoPolice provides a foundation for developing models capable of identifying events of interest in large-scale body-worn camera video repositories, enabling more efficient downstream human review.

Large language models (LLMs) increasingly issue judgments read as binary verdicts, and a growing literature reports such judgments shifting under logically irrelevant changes of wording - among them an amplified yes-no bias on moral dilemmas, absent in humans. A single framing cannot say what such a shift is: in a yes/no question the word "no" is at once logical verdict, lexical token, and last-printed option. We introduce a psychometric battery that separates these: crossed symmetrization - every logically irrelevant factor flipped in balanced pairs - across a corpus of question forms. A graded rating across logically equivalent forms recovers a coherent internal moral scale: frontier models' stance $θ$ is nearly format-invariant (cross-form incoherence 0.12-0.21 on a $\pm 1$ axis); small open-weight models fail in model-specific ways. Forcing the verdict through yes/no overlays a decomposable artifact: an order bias toward the last-printed option - opposite to classic human primacy - plus a lexical pull toward the word "no"; the artifact is substantial only in the Claude models (story-averaged -0.32 to -0.86), $\approx 0$ for GPT-5.5 and Gemini, and shrinks under extended reasoning. The word and the verdict share one token; swapping the words for arbitrary labels separates them, and the verdict-attached logical bias proves $\approx 0$ for every frontier model, while model-specific label and order attachments remain: the models are not drawn toward rejecting - the pull follows the printed surface, not the verdict it carries. A minimal model, $P = σ((θ\pm m)/s)$, summarizes any such artifact by a framing susceptibility m and a moral decisiveness s, measurably distinct from sampling temperature. The battery applies unchanged to any dilemma set and binary format: measuring what a model values requires crossing the frames of the question, not asking once.

While humans develop core visual skills long before acquiring language, contemporary Multimodal LLMs (MLLMs) still rely heavily on linguistic priors to compensate for their fragile visual understanding. We uncovered a crucial fact: state-of-the-art MLLMs consistently fail on basic visual tasks that humans, even 3-year-olds, can solve effortlessly. To systematically investigate this gap, we introduce BabyVision, a benchmark designed to assess core visual abilities independent of linguistic knowledge for MLLMs. BabyVision spans a wide range of tasks, with 388 items divided into 22 subclasses across four key categories. Empirical results and human evaluation reveal that leading MLLMs perform significantly below human baselines. Gemini3-Pro-Preview scores 49.7, lagging behind 6-year-old humans and falling well behind the average adult score of 94.1. These results show despite excelling in knowledge-heavy evaluations, current MLLMs still lack fundamental visual primitives. Progress in BabyVision represents a step toward human-level visual perception and reasoning capabilities. We also explore solving visual reasoning with generation models by proposing BabyVision-Gen and automatic evaluation toolkit. Our code and benchmark data are released at https://github.com/UniPat-AI/BabyVision for reproduction.

We present Sparse Backdoor, a supply-chain attack that plants a provably undetectable backdoor in pre-trained image classifiers, including convolutional networks and Vision Transformers. The attack injects a structured sparse perturbation along a randomly chosen direction into a small subset of columns at each fully connected layer, propagating a trigger signal to an adversary-chosen target class, and masks the perturbation with an independent isotropic Gaussian dither. The dither serves a single technical purpose: it induces a clean reference distribution anchored at the pre-trained weights, against which undetectability can be formalized. Under a mild margin condition on the pre-trained classifier, we show that the dithered reference is functionally equivalent to the original classifier. We prove that distinguishing the backdoor-injected model from this reference is at least as hard as Sparse PCA detection, which is computationally infeasible under standard hardness assumptions. The guarantee holds against any probabilistic polynomial-time distinguisher with white-box access to the parameters.

Camera intrinsics are vital for recovering 3D structure from 2D video. However, most 3D algorithms assume fixed intrinsics throughout a video, an assumption that often fails for real-world in-the-wild videos. Consequently, estimating per-frame intrinsics from RGB images is critical for making 3D methods robust to videos with dynamic intrinsics. InFlux previously advanced this research direction by establishing the first real-world benchmark with per-frame ground truth intrinsics for dynamic intrinsics videos. Nevertheless, existing methods remain inaccurate due to two obstacles: (i) training data is scarce and lacks intrinsics diversity; and (ii) benchmarks, including InFlux, have limited scene and camera motion diversity, making it difficult to properly evaluate methods. To address both gaps, we present InFlux++, consisting of two components. InFlux++ Synth is a large-scale procedurally generated synthetic video dataset with 441K+ annotated frames from 1841 high-resolution videos, providing accurate per-frame ground truth intrinsics for training dynamic intrinsics prediction models; a subset also includes per-frame pose, depth, and normals. The videos feature rich intrinsics diversity through changes in camera zoom and focus, as well as dynamic objects and realistic rendering effects such as lens distortion and defocus blur. InFlux++ Real is a large-scale real-world benchmark that extends InFlux with 514K+ newly captured frames across 334 high-resolution videos, spanning a wider range of scenes and camera motions. Finetuning existing intrinsics prediction methods on InFlux++ Synth consistently improves focal length estimation across both InFlux++ Real and InFlux, suggesting that synthetic supervision is promising for RGB-based intrinsics prediction. For the dataset, benchmark, code, videos, submission instructions, and live leaderboard, please visit https://influx.cs.princeton.edu/ .

Self-distillation is a promising recipe for self-improvement in language models. In this setting, a model can serve as its own teacher when given privileged information, such as a solution to a math problem. This seems especially appealing for thinking models, which can use test-time reasoning to absorb the privileged information. Surprisingly, we show that privileged self-distillation degrades thinking models on long reasoning traces: across five Qwen3 and OLMo thinking models evaluated on AIME24, AIME25, and HMMT25, privileged-context distillation causes a relative drop of up to 17% in avg@16 accuracy. The degradation scales with the amount of privileged context withheld from the student and is most pronounced at long rollout budgets, where thinking models otherwise obtain their largest gains. This failure mode is not specific to self-distillation: on-policy distillation (OPD) improves thinking models, but privileged OPD reverses these gains. Our diagnostics link this failure mode to how privileged teacher context reshapes learning at high-entropy forking positions, where multiple continuations remain plausible and may lead to different reasoning paths. Privileged context lowers fork rates in thinking-model rollouts but not in instruction-model rollouts. This leads to an interesting dichotomy, where privileged context can help instruction-tuned models but hurts stronger thinking models. The effect is visible when the student begins a self-correction branch, where privileged OPD penalizes sampled reconsideration tokens that vanilla OPD supports. Thinking models trained with a privileged teacher produce fewer verification, backtracking, and hedging markers, even after length normalization. These findings indicate that self-distillation for strong thinking models requires attention to token-level signal, especially around correction and reasoning steps.

The performance of deep learning models crucially depends on the settings of hyperparameters like learning rate, initialization scale, and weight decay. Hyperparameter transfer aims to make near-optimal hyperparameter settings consistent across model scale, so that large models can be optimized by proxy tuning their smaller, cheaper-to-optimize counterparts. While transfer principles are well-studied in the context of dense neural networks in language and vision tasks, they remain comparatively under-explored for graph neural networks (GNNs). We develop and validate a transfer parameterization for GNNs trained with SGD, Adam, and AdamW. Through theoretical scaling analyses and controlled experiments, we show that the proposed parameterization yields stable feature updates, learning rate transfer, and improved performance as width and depth increase. For SGD, we identify graph-dependent first-layer correction factors and show that their use can accelerate early training in graphs with sparse bag-of-words inputs. For Adam, we explore how different message passing normalizations affect early- and late-training transfer behavior, illustrating the importance of message passing normalization and advocating for an associated hyperparameter. For AdamW, we adapt a parameterization that allows for the joint transfer of weight decay and learning rate. Together, these results provide a practical recipe for scaling GNNs across a variety of learning tasks and training scenarios.

Extending automatic speech recognition (ASR) to low-resource African languages is constrained by the prohibitive demands of data collection at scale. A promising direction is to leverage linguistic relatedness to enhance cross-lingual transfer from a related auxiliary language to the low-resource target by sequentially adapting on both. Although this strategy has shown meaningful improvements in small ASR models, its effectiveness in large ASR remains unclear. We extend this framework to large multilingual ASR through a systematic controlled experimental design spanning six factors, two Africa-centric corpora, and four large ASR models, isolating whether linguistic relatedness reliably predicts cross-lingual transfer gains in this setting. Across all conditions, pre-adaptation on related auxiliary languages yields no practically meaningful transfer improvements given minimal target-language data, suggesting that linguistic relatedness alone may not reliably predict cross-lingual transfer gains in large multilingual ASR, or constitute an effective strategy for extending such models to low-resource languages.

Diverse applications for robotics, such as industry and agriculture, require robots to operate across various embodiments, changing visual conditions, and complex planning. Vision-Language Models (VLMs) offer a promising foundation for general-purpose and interpretable robotic reasoning. Aligning VLMs with diverse robot applications requires a modular understanding of the individual decision components that underlie robotic behavior. Capturing such structure is challenging for conventional robot benchmarks that are primarily based on teleoperated, end-to-end datasets. We propose Robot Question Answering (RQA), a modular evaluation framework and RoboVista, a benchmark curated from real robotic systems, research papers, and expert annotations. RoboVista contains 474 Visual Question Answering (VQA) instances with human annotated reasoning and covers 39 unique task types in agricultural, industrial, domestic, surgical robotics, autonomous driving, and open robot datasets. Experiments on RoboVista show that state-of-the-art VLMs exhibit substantial gaps. Physical robot experiments suggest strong correlation between RoboVista performance and real-world task execution.

Motivated by the challenge of stabilizing a general unknown linear dynamical system (LDS) from observations, we study the natural prerequisite of online prediction. Our goal is to achieve sublinear regret with a memory footprint that adapts to the intrinsic complexity of the dynamics rather than the full hidden-state dimension. We focus on the practically central regime of systems with low instability complexity -- eigenvalues outside the real stable interval that do not decay rapidly, together with non-semisimple modes -- potentially embedded in an otherwise stable real spectrum of much higher dimension; we write $k$ for this count. This regime is the primary setting in which stabilization is plausible: we show that many systems with high instability complexity cannot be stabilized without exponentially large controls. Thus, prediction is meaningful for stabilization precisely when the instability complexity is small. Within this regime, we introduce a unified online algorithm that handles every LDS (including non-diagonalizable systems with complex or exploding modes) with a learnable parameter count of $\widetilde{O}(k)$. Finally, we prove a lower bound showing that $k$ is a valid complexity measure: any filter-based predictor needs at least $k$ filters. Experiments corroborate our theory: on a high-dimensional system, our predictor sharply outperforms prior methods at an equal parameter budget.

Modern AI progress has been driven by ML methods that are generalizable across settings and scalable to larger regimes. As large language models demonstrate advanced capabilities in reasoning, coding, and engineering tasks, it is increasingly important to understand whether they can discover such methods rather than only apply existing ones. We introduce MLS-Bench, a benchmark for evaluating whether AI systems can invent generalizable and scalable ML methods. MLS-Bench contains 140 tasks across 12 domains, each requiring an agent to improve one targeted component of an ML system or algorithm and demonstrate that the improvement generalizes across controlled settings and scales. We find that current agents remain far from reliably surpassing human-designed methods, and that engineering-style tuning is easier for them than genuine method invention. We further study the effects of test-time scaling, adaptive compute allocation, and context provision on agents' discovery performance, together with case studies of their behavior. Our analyses suggest that the bottleneck is not only in proposing new methods, but also in the scientific insight needed to plan, validate, and scale claims about them. More search, compute, or context alone does not remove this bottleneck. We build and maintain a community platform for cumulative and comparable iteration, and release the data and code at https://mls-bench.com.