How is language identity represented and causally maintained in transformers, and to what extent do the mechanisms identified in one language set transfer to other languages and training regimes? We introduce Language Identity Head Ablation (LIHA), a causal intervention that zeros each attention head individually and measures the resulting language switch rate. Across our multilingual evaluation sets, LIHA identifies heads whose ablation has disproportionate effects on output language. In GPT-2, a small set of first-token broadcaster heads---led by L6H1 (switch rate 0.32, 3.23$\sigma$ above the population mean)---attend persistently to the first prompt token across query positions. The combination of persistent first-token attention and causal sensitivity suggests that these heads participate in maintaining output language. Following ablation, first-token attention increases in heads above the ablated layer for the ablations tested (permutation test against random five-head sets, $p < 10^{-5}$). We do not test whether these heads functionally compensate for the ablated head. We also compare Qwen2.5-1.5B-Base and Qwen2.5-1.5B-Instruct, which are matched in architecture and parameter count but differ in training history. The base model is nearly flat (max SR$=$0.016, 200/336 heads at SR$=$0.0), whereas the instruct model concentrates causal influence at layer~0, led by L0H5 (SR$=$0.224, 8.93$\sigma$ above the mean). Finally, GPT-2 heads identified from the European-language experiments show no observed switch effects on the tested Chinese and Russian prompts, while other early-layer heads do show effects. Because script, typology, and tokenization vary together in this comparison, these results do not isolate a specifically script-driven mechanism. Repository is available at https://github.com/ajlrza/First-Token-Broadcasters.
Self-supervised learning (SSL) typically relies on a backbone encoder followed by a small multilayer perceptron (MLP) projection head, which is conventionally discarded after training, while backbone quality is assessed via costly linear probing on labeled data. We argue that this approach including discarding the projector is rather computationally wasteful. Instead, we propose replacing the MLP head with a radial basis function network (RBFN), whose interpretable center and shape parameters can be exploited to judge representation quality without labels or a separate classifier. To this end, we introduce Scale-Normalized Separation (SNS), a novel label-free quality metric derived solely from the kernel centers and shapes learned during training. Across five canonical SSL architectures (MoCo, SimCLR, BYOL, SwAV and SimSiam) and four image classification datasets, we show that RBFN projection heads are competitive drop-in replacements for standard MLP projectors. We recommend constructing them with three RBF layers activated by the Gaussian radial basis function. Moreover, SNS exhibits strong to very strong positive correlation with established logistic regression metrics, demonstrating that a trained RBFN projector can act as a reliable proxy for backbone representation quality. We additionally publish a novel PyTorch compatible image classification dataset based on Google's Open Images V7 to facilitate reproducible research into representation learning.
We introduce WildBox, a dataset and benchmark for monocular 3D detection of wildlife from drone video, comprising 237,505 3D bounding box annotations across seven African savanna species grouped into six benchmark classes. Annotations follow a KITTI/Omni3D-compatible format in a per-segment scale-normalised camera frame, with instance identities maintained across each segment. We evaluate two open-vocabulary monocular 3D architectures, OVMono3D-LIFT and DetAny3D, under zero-shot, ground-truth 2D box prompt, and supervised fine-tuning protocols. Open-vocabulary 2D foundation models provide usable zero-shot wildlife localisation (50.55 AP@50), but zero-shot 3D detection collapses to 0.00 AP across both architectures and every 2D-input condition tested, including ground-truth 2D box prompts, thus isolating the failure to the 3D stage. Fine-tuning on WildBox recovers performance to 8.68 +/- 0.47 AP-BEV@0.50 and 13.17 +/- 0.69 AP3D macro. Depth contributes 84% of normalised Hausdorff distance after fine-tuning and over 99% in zero-shot, identifying monocular aerial depth as the dominant open problem in this regime. A coarse-to-fine curriculum, i.e. pretraining on a merged zebra class before fine-tuning on the Grevy's/plains split, improves macro 3D performance with less total compute, with the largest gains on the two zebra subclasses. WildBox is released with video-level splits and baseline checkpoints to enable progress in 3D wildlife perception from drone video.
Vision-Language Models (VLMs) often struggle with strategic reasoning in a multi-agent setting, i.e., deciding under uncertainty while coordinating with and adapting to other agents. Real-time strategy (RTS) games can be a natural testbed for this capability, requiring long-horizon planning, coordination with allies, and adaptation to opponents under partial observability. But existing RTS benchmarks cover limited agent configurations, lack controlled evaluation across distinct strategic demands, and remain confined to fixed scenarios. To address these gaps, we propose RTSGAMEBENCH, built on Beyond All Reason, an RTS game whose larger scale broadens the strategic space beyond prior testbeds. In RTSGAMEBENCH, we offer holistic evaluation across diverse full-game matchups, diagnostic mini-games each emphasizing a focal strategic demand, and extensible coverage via a self-evolving framework that turns free-form queries into new mini-games, improving with accumulated experience. We further propose RTSGAMEAGENT, an agent harness that combines scalable group-level trajectory planning with short- and long-term memory and outperforms existing agents. We find that state-of-the-art VLMs struggle where humans succeed, particularly when they must coordinate with allies or contend with multiple opponents.
Dataset distillation (DD) has emerged as a prominent approach in data centric machine learning, aiming to synthesize compact training sets for efficient training by compressing the information in large datasets into a small number of synthetic samples. However, DD methods are often evaluated under inconsistent evaluation protocols, ranging from standard ERM to single/multi-teacher supervision, making it difficult to isolate the effectiveness of distilled data from evaluation. Moreover, many prior methods claim that DD outperforms data pruning approaches such as coreset selection (CS), based on the assumption that restricting condensed datasets to subsets of real samples fundamentally limits their expressiveness. In this work, we critically evaluate DD methods through large-scale experiments using standardized datasets and evaluation protocols to assess their intrinsic effectiveness. We benchmark seven state-of-the-art (SOTA) DD methods on ImageNet-1K, ImageNet100, and ImageNette, using three widely adopted training protocols against four CS strategies. Our results show that while some DD methods fail to outperform even simple random subsets, the SOTA DD approaches are comparable to or worse than coresets on large-scale datasets and incur a substantially higher cost for construction. Beyond accuracy, we also evaluate the representativeness, diversity, and quality of condensed sets, and find that coresets consistently achieve better coverage of the original data distribution. These findings highlight the limited practical advantages of current DD methods and show that coresets remain competitive and are often a more computationally efficient alternative for data-centric learning.
Benchmark scores tell you what an agent got right; they do not tell you how it got there. In this work, we introduce methods for comparing agents procedurally in different contexts, where the model, tasks, and approaches vary. We compare ten agents and find that they are identifiable by their behavioral habits, which we define as fingerprints: a probe over these procedural signatures attributes an unseen trajectory to the correct agent at 85.7% accuracy. We develop procedural representations for agent problem-solving procedures with an emergent vocabulary induction technique which is meant to be maximally compressive to avoid surface-level variation while being expressive enough to unveil the quirks of the models' patterns. We apply our framework to the software engineering evaluation dataset SWE-Bench to study the structural distinctness of agent trajectories and find that behavior is most similar between models from similar release periods and those that are distilled from one another (e.g., a distilled student model and its teacher have a Jensen-Shannon divergence of 0.25, about half the distance between other model pairs). As more models saturate evaluations, we believe that it will be important to probe model behavior along more holistic dimensions than success rates alone. We introduce ProcGrep, a library for auditing and evaluating agents for how they approach tasks at a procedural level given their traces in a top-down fashion. ProcGrep is available at https://github.com/hamidahoderinwale/procgrep.
Digital twins are often built under the assumption that greater model fidelity leads to better downstream performance. We show that this is not always the case: retaining weakly dependent or auxiliary degrees of freedom can increase the dimensionality of learning and control problems enough that a sparse stochastic surrogate outperforms even an exact high-dimensional model. To identify such structure, we introduce a framework for dependence discovery directly at the level of the full conditional law. The method combines conditional generative modeling with Gaussian-process distillation and kernel mode decomposition to quantify how strongly individual variables enter the conditional distribution, including through variability, tails, multimodality, and other features invisible to the conditional mean. The resulting dependence structure can be used both to construct sparse stochastic digital twins from data and to reduce known high-dimensional models for downstream tasks. In a modified Lunar Lander problem, an eight-dimensional sparse surrogate yields better control performance than the exact 109-dimensional simulator despite lower transition fidelity. A stochastic heat-equation example shows that the framework recovers sparse structure in a 102-dimensional data-driven setting and produces a more effective surrogate than one conditioned on all candidate inputs, while economic data demonstrate its use for identifying time-varying memory structure. These results show that model fidelity and downstream utility are distinct objectives, and that dependence discovery provides a practical route to sparse stochastic digital twins.
Robust deep learning under heavy-tailed and impulsive noise remains challenging because conventional losses such as mean squared error (MSE) exhibit unbounded sensitivity to outliers. Although correntropy-based objectives improve robustness, many conventional formulations rely on fixed kernel parameters that must be empirically tuned and remain static during training. To address these limitations, we propose an Adaptive Log-Correntropy Loss (ALCL), a heavy-tailed loss formulation that adaptively learns its robustness geometry during optimization. ALCL introduces a logarithmic residual model whose shape and scale parameters are learned jointly with network weights through differentiable reparameterization. This yields a heavy-tailed robust optimization objective whose psi-function is formally bounded and redescending, allowing the loss geometry to adapt dynamically to evolving residual statistics while suppressing extreme outliers. Comparative experiments on four widely used benchmark datasets spanning grayscale and red-green-blue (RGB) image data under mixed heavy-tailed and impulsive noise demonstrate that ALCL consistently outperforms MSE and optimally tuned generalized correntropy losses in both reconstruction fidelity and downstream classification accuracy. While performance differences remain small under low-noise conditions, under high-noise regimes ALCL improves median accuracy by up to 4.75% on grayscale benchmarks and 4.51% on RGB datasets, with reduced variance across runs. These results demonstrate that adaptive robustness through joint learning of loss parameters provides a computationally efficient alternative to static correntropy-based losses for deep learning in non-Gaussian environments.
We study the task of CVE-conditioned exploit generation, where a model drafts proof-of-concept (PoC) exploits given software vulnerability context. We adopt a data-centric approach, constructing a high-quality dataset via multi-stage preprocessing and introducing a scalable evaluation framework with LLM-as-judge and fine-grained rubrics. Under this unified setup, we benchmark 17 large language models across 8 evaluation criteria, providing systematic insights into their zero-shot capabilities. We further show that a compact 8B open-weight model, when fine-tuned on curated data, achieves over 42.5% improvement in exploit quality and rivals some proprietary models when combined with simple test-time rejection strategies. Our results highlight the importance of data quality, structured supervision, and evaluation design for reliable exploit generation, suggesting that these factors can be as critical as model scale in adapting LLMs to cybersecurity tasks.
Bimanual robot systems substantially expand manipulation capabilities, but coordinating two arms introduces additional control complexity and failure modes that are not well captured by existing benchmarks. We introduce DuoBench, an extensible benchmarking framework for bimanual manipulation policies on the FR3 Duo platform. DuoBench comprises eleven tasks spanning four coordination categories, implemented in simulation and partially replicated in the real world through reproducible task recipes with 3D-printable assets. In addition, we propose a stage-based evaluation scheme that supports fine-grained semantic failure analysis beyond binary success. It includes stage-conditional pass rates with confidence intervals as well as human-teleoperated datasets for all benchmark tasks. We benchmark several imitation-learning and vision-language-action policies in simulation and on real hardware. Our results show that current policies remain challenged by bimanual manipulation, particularly at the beginning of interaction, during parallel arm execution, and in sim-to-real transfer settings. We further find that simulation reveals the same failure modes as real-world experiments. DuoBench provides a reproducible testbed for analyzing them and studying future methods for dual-arm policy learning. Code, datasets, and videos are available at https://duobench.github.io
Spoken factual claims often occur within multi-turn conversations, where surrounding dialogue can provide context unavailable from the claim alone. Yet most fact-checking research evaluates isolated text, leaving conversational audio under-studied. We introduce MAD2, a synthetic Multi-turn Audio Dialogues benchmark for spoken claim verification with 1,000 two-speaker dialogues, 1,230 sentence-level check-worthy candidate annotations, and approximately 10 hours of audio. We also propose calibrated multimodal fusion of a context-aware audio encoder and a dialogue-aware text model. Adding dialogue context improves verification across settings, although the gains vary by scenario. Past-only context often approaches local offline performance, suggesting its usefulness when future context is unavailable. Fusion achieves its highest mean advantage over text with full-dialogue context, but does not consistently or significantly outperform text across settings. Under full-dialogue context, exploratory subgroup results show greater performance variation across dialogue scenarios for text and fusion, but across spread styles for audio.
Purchased artificial intelligence (AI) inference belongs in Scope 3 Category 1 of a corporate greenhouse gas inventory, yet practice omits it or applies a spend-based factor that gives 1.7 to 12.3 times the physical estimate and cannot tell one model or region from another. This paper sets out a token-based method and open dataset: a cradle-to-gate boundary, four tiers of activity data, an effective-token basis, a four-term factor (serving electricity, embodied hardware, data-centre construction, disclosed final training runs) and a rule resolving any model identifier and region to a factor. Inference footprints span twelve times between model classes and nine between regions. The mid-class factor, 0.039 kg of carbon dioxide equivalent (CO2e) per million billed tokens at the world-average grid, is the low end of a stated structural range: a less efficient serving configuration raises it 2.8 times, and replacing the training term with all model development, on its own boundary, gives 1.5 to 2.9 times it. Input bands alone give -18% to +29%. Published per-query energy estimates that can describe a mid-class model are 2.4 to 7.2 times the mid-class energy, above its structural range in four of five rows; those that can describe a frontier-class model bracket it at 0.7 to 1.6 times. Worked examples give 8.5 kg CO2e a year for a 200-person firm, and 2.1 tonnes for fifty developers using agentic coding tools on a mid-class model or 8.3 tonnes on a frontier-class one. The contribution is a traceable figure; accuracy against metered energy is untested.
Compositional priors describe the generic properties of layered functions in deep Bayesian models, where deep neural networks with random weights are a canonical example. In the wide-network limit, the prior is a Gaussian process (GP) with a depth-dependent kernel, and its behaviour as depth grows has been extensively studied through this kernel. Here, we study another case, where each layer itself is a vector valued Gaussian process, and our aim is similarly to understand the limiting behaviour of the prior as depth grows. Previous GP work has established that for the radial basis function (RBF) kernel and a certain range of bandwidths $r$, the prior degenerates in the limit, converging to the set of constant functions, which is not useful as a probabilistic model. In this paper we establish several new results. First, we identify a sharp bandwidth threshold $r_c(d) = \Theta(\sqrt{d})$ above which the limit is degenerate, strengthening the earlier bounds. Second, and more importantly, we show that for $r$ below the threshold $r_c(d)$ the prior converges to a limit distribution $\pi_{\bar{Z}}$. We also prove that these distributions are non-degenerate and non-Gaussian, with non-vanishing dependence between coordinates. In contrast to the previously known degenerate regime, deep Gaussian process priors can therefore admit non-trivial limits. Empirically, we verify the threshold across a range of dimensions $d$, and demonstrate a complex multimodal behaviour of the limit distributions $\pi_{\bar{Z}}$, a regime that becomes increasingly narrow with $d$ and would be hard to identify without knowing the threshold.
The rapid development of large language models(LLMs) has led to remarkable advances in natural language processing. However, the increasing scale of these models introduces substantial challenges in terms of storage, transmission, and deployment. Though great efforts have been devoted to model compression and quantization, existing methods often rely on fine-tuning or calibration data, which exhibit limited generalization across different tensor types. In this paper, we argue that video codecs offer a promising solution for LLM compression, due to their inherent compatibility with matrix structured data, configurable compression strategies, and the availability of highly optimized, off-the-shelf implementations. Therefore, we present LLMCodec, a video codec-based LLM compression method that integrates affine quantization with the recent VVC/H.266 video codec. Beyond VVC, we further compare a range of video codecs and encoding profiles to evaluate their impact on compression performance. Experiments on different models demonstrate the robustness and generality of LLMCodec. Notably, on LLaMA-3-8B at 2-bit precision, LLMCodec reduces perplexity by over 1.5x and improves downstream task accuracy by 21% compared with the existing method.
Reconstructing interactive, simulation-ready 3D scenes from a single image is a critical bottleneck for robotic manipulation. While recent single-image lifters recover plausible per-object shapes, composing them yields scenes that collapse under physical simulation due to interpenetrating, hovering, or sinking objects. Existing physics-aware compositional reconstruction methods mainly address this as a post-hoc layout correction, leaving the underlying geometric errors unresolved. To address this, we introduce SimuScene, a compositional 3D reconstruction pipeline that puts physics in the loop of shape and layout estimation. Rather than using physics merely for layout cleanup, we utilize the physics engine as a diagnostic measurement tool during the generative process itself. By diagnostically simulating reconstructed objects under gravity, we convert penetration and support failures into quantitative correction signals that drive amodal shape resampling, with gravity-axis stretching as a lightweight correction for minor errors. This physics-informed feedback loop mitigates accumulated reconstruction errors and produces a stable, simulation-ready compositional 3D scene. Extensive experiments demonstrate state-of-the-art performance on physical stability and geometric alignment benchmarks. We further highlight SimuScene's utility by deploying reconstructed environments in humanoid control and robot-arm manipulation tasks.
As older adults (OAs) increasingly favor aging in place, remote family members (RFMs) need meaningful awareness of their loved ones without intrusive monitoring. Large language models can provide rich interpretations of non-intrusive sensing data, yet recent work has largely emphasized technical accuracy over how families use or benefit from this information. We conducted formative interviews with 12 RFMs and 2 OAs, followed by a three-month deployment case study with one OA-RFM dyad. Our system provided daily AI-generated summaries of multimodal sensing data, which we iteratively refined with both parties using a research-through-design approach. Although the summaries integrated well into the dyad's daily routine, sustaining over 90% engagement, their lived experience revealed evolving tensions around agency, reassurance, privacy, trust, and communication, shaped by sensor limitations and existing communication misalignments. Our findings inform the design of future AI-mediated sharing systems that account for real-world sensing constraints and complex family care dynamics.
We study how switching costs affect minimax regret under partial feedback. Which features of the feedback preserve the $\sqrt T$ rate, and which force the $T^{2/3}$ rate? We address this question in the framework of feedback graphs, which interpolate between full information and bandits: playing an action reveals the losses of its out-neighbors in a known directed graph. Prior work on feedback graphs with switching costs gave algorithms with regret of order $T^{2/3}$, but matching lower bounds only for restricted classes of graphs, leaving the question open. We establish a complete characterization of the minimax regret with switching costs for all feedback graphs: $\widetilde\Theta(\sqrt T)$ if the graph is strongly revealing, meaning that every pair of distinct actions has an endpoint that observes the losses of both; $\widetilde\Theta(\gamma^{1/3}T^{2/3})$ for every other observable graph, where $\gamma$ is the domination number; and linear otherwise. In particular, tournaments, in which every action observes its own loss and each pair of actions is joined by one directed edge, sit at the phase transition: they behave like full information, but deleting the edge of any pair forces the rate $T^{2/3}$. Hence an action that observes all losses is neither necessary nor sufficient for the rate $\sqrt T$. Our algorithm TournamentSteering for the strongly revealing case rests on a new principle: each observed comparison of two actions serves both to estimate losses and to steer play, and the loss reduction from steering pays for the variance of the estimates. For every strongly revealing graph, distributing the comparisons among observing actions keeps this variance, and hence the regret, polylogarithmic in the number of actions, by an argument similar to the greedy construction of logarithmic dominating sets in tournaments.
Distilling expert demonstration data into large generative models using behavioral cloning is a scalable approach to learning capable policies for robotic control, particularly for dexterous manipulation. Reinforcement learning (RL) can be used as a means to finetune these policies further using additional experience. An open question is whether RL is more sample-efficient than collecting more expert demonstrations. Prior work has finetuned large pretrained policies in a scalable fashion by applying RL to a smaller residual policy that corrects the pretrained model. However, for typical sparse reward tasks, RL algorithms can struggle to optimize the behavior in a sample-efficient manner. We explore inverse reinforcement learning, where a dense reward function is learned from expert demonstrations, potentially reducing the challenge of RL finetuning. We specifically consider coherent imitation learning, an IRL method that facilitates improvement of the BC policy by using a specific learned reward formulation with theoretical guarantees for imitation learning. We show that our IRL method maintains or improves the performance of pi-0.5 on all six sparse manipulation tasks and achieves a $\geq 90\%$ success rate on five out of six complex manipulation tasks, outperforming RL-based baselines using sparse rewards. We also demonstrate real-world learning on a robot manipulation task, finetuning pi-0.5 to success with less than 30 minutes of environment interaction time. By ensuring our initial pretrained finetuning policy is optimal for our initial reward and critic, our method circumvents the initial drop commonly seen in RL finetuning and enables faster improvement.
The goal of Sign Language Production (SLP) is to translate natural language text into corresponding sign language movements performed by a virtual avatar. The quality of the generated motion is typically evaluated by a motion-space Frechet distance (FID) and back-translation BLEU score on benchmarks such as How2Sign. The existing approaches can improve both metrics substantially while the underlying generator fails to represent the input sentence, a failure we call conditional collapse. In this work we propose to evaluate the generated motion at three independent levels: (tau1) initial-pose conditioning, (tau2) output diversity, and (tau3) target faithfulness. We compute these as pairwise-distance ratios using latent representations of a frozen motion autoencoder (MoAE), judged against each dataset's own GT-vs-GT reference. We evaluate diffusion and regression SLP models on How2Sign, including a re-implemented Neural Sign Actors, and show that tau3 faithfulness is never attained, while FID varies by nearly two orders of magnitude and is uncorrelated with faithfulness. We then apply the same models and diagnostic to Phoenix-2014T, CSL-Daily, and gloss-level ASL3DWord and show that faithfulness varies proportionally to how densely the training sentences cover the test conditions (conditioning coverage) and not corpus size. The smaller but densely covered Phoenix-2014T and ASL3DWord datasets generate faithful sign motion, while CSL-Daily follows the input text yet falls short of sign language faithfulness. Using the largest, open-domain How2Sign dataset, the models collapse at every scale. We introduce two measurements of the text data that rank the corpora in the order of how densely the training sentences fill the embedding space, and how close each test sentence lies to its nearest training sentence.
Generative modeling and self-supervised representation learning (SSL) optimize structurally different objectives: generative training rewards distributional fidelity, while SSL rewards semantic coherence. Yet recent work repeatedly finds that SSL features improve generative training, though the mechanism of this synergy remains unclear. Here, we study the benefits of SSL in generative modeling in the framework of one-step generation where the role of representation is explicit: frozen SSL features are used to match generated samples to real data. We use the Sinkhorn divergence in that feature space, providing a tractable surrogate for the Wasserstein distance, the population-level discrepancy approximated by Fr\'echet-style evaluation metrics (such as FID). We find that this objective becomes highly effective when computed in a semantically structured SSL feature space (a 39$\times$ reduction in ImageNet FID). We trace this behavior primarily to matching estimation: semantic SSL features that suppress nuisance reconstruction details induce a more compact geometry, making distribution matching more tractable. As a consequence, the best training SSL features need not match the features used by the evaluation metric. In particular, we show that using Inception as the feature extractor can improve FID while degrading matching stability and sample quality, revealing a form of metric hacking. Using extensive experiments on ImageNet, we identify which SSL feature families lead to best generation performance and show that matching stability is a quantitative criterion for selecting them. Code is available at https://github.com/Genentech/semantic-transport-generation.