We study whether \emph{long-horizon model-based value expansion} can serve as an effective framework in pixel-based offline reinforcement learning (RL). Long-horizon model-based value expansion from pixels is substantially harder than state-based environments, as the agent must learn both the representation and its dynamics, and errors from either can accumulate recursively over imagined trajectories. We introduce MAC-v2, which enables long-horizon model-based expansion from pixels by training a reconstruction-free latent world model with action chunking. By learning representations from prediction and control objectives, MAC-v2 does not need to encode control-irrelevant visual details that can hinder the dynamics prediction. Action chunking further mitigates the prediction error by reducing the number of recursive model predictions required for value expansion. Across diverse visual benchmarks including visual OGBench, MMBench2, Robomimic, and LIBERO-10, MAC-v2 achieves strong performance against model-free and model-based offline RL baselines. Our analyses show that long-horizon model-based value expansion is critical for the performance of MAC-v2, and that both action chunking and latent learning objectives are essential for reliable long-horizon latent imagination.
On-policy distillation (OPD) trains a student model to match a teacher's next-token distributions on student-generated prefixes. Extrapolative OPD further extends the teacher's log-probability shift relative to a reference policy to construct stronger learning targets. However, existing methods couple extrapolation direction and distance, allowing the magnitude of the teacher's shift to dictate how far the target moves. We propose $D^2$-OPD, decoupling these two factors: a per-prefix hard KL constraint bounds the target's distance from the teacher, while an adaptive combination of the teacher's post-training shift and the initial student--teacher gap determines the extrapolation direction. We optimize the student toward this target using forward KL, providing learning signals even for low-probability tokens. We evaluate three settings: where standard OPD performs well, where it struggles with base-model initialization, and where no teacher reference is available. On mathematical reasoning benchmarks, $D^2$-OPD attains the highest score in all three settings, outperforming the strongest baseline by up to 9.45 points. Ablations validate the contributions of each component.
AI agents may consume massive tokens across a wide range of tasks, but how many of those tokens are really necessary? Although prior work has shown that agents can incur unexpectedly high token costs, we still lack a systematic way to identify and measure token waste. In this work, we introduce a unified framework for measuring token waste across heterogeneous agent systems. Our framework records and normalizes complete agent trajectories, identifies removable spans using an LLM-as-a-Judge (LLMaJ), and maps the identified text spans back to the original trajectory for token-level measurement. We validate the identified waste through direct interventions by removing marked spans, resuming agent execution, and measuring changes in task outcomes and subsequent actions. In online pruning experiments, we remove 41-46% of the model's reasoning while largely preserving task performance across two distinct agent benchmarks. Using this framework, we analyze agents across multiple settings. Our analysis shows that token waste has a system-level structure that aggregate token usage alone cannot reveal. First, waste varies more across tasks and agent settings than across models. Second, reasoning and visible execution can exhibit different waste patterns. Third, different harnesses can introduce distinct sources of inefficiency even under the same model and task. Overall, our framework transforms token waste into an actionable optimization signal, enabling researchers and practitioners to train more efficient models, design better reasoning strategies, and optimize agent harnesses toward reducing unnecessary computation rather than merely reducing token usage.
Neural networks admit many parameterizations of the same or nearly the same function, but deployment typically stores whichever parameter vector training happens to produce. We show that this representational freedom can itself be used for compression. Rather than redesigning the codec, we search over exact function-preserving transformations and controlled functional perturbations, using the unchanged physical compressor itself to select cheaper representations. Exact gauge families provide a zero-distortion reference under the strict audit; allowing small functional changes exposes much larger rate-distortion gains. Across SmolLM2 and Qwen3, our method reduces Zstandard-compressed checkpoint size by up to 23.4\%, with similar relative reductions under gzip, Brotli, and ZipNN, establishing trained-network representational freedom as a practical storage resource.
At context lengths of about 1M tokens, dense attention is costly in prefill and decoding, and training typically relies on long documents, which become scarcer as the target length increases. Most sparse attention methods score all preceding blocks or tokens, so their selection cost per query is $O(L)$ at context length $L$; hierarchical selection reduces it to $O(\log L)$, but prior trainable hierarchical methods report accuracy only within their training length. We propose RecSA (Recursive Sparse Attention), which builds a tree over the KV cache in which each node summarizes its subtree with one key and one mean-pooled value. The keys come from one pooler per layer, shared by all levels and distilled from dense attention. Each level above the tokens selects at most $k$ candidates, so each query has $O(\log L)$ selection cost and reads $O(\log L)$ entries; the summaries of unselected candidates share one softmax with the exact tokens. Each entry's RoPE position is its index among the candidates of its level, so at any length every relative distance stays within the range seen during training. Initialized from Qwen3-4B and trained on 0.62B tokens at the original model's 32K training length, RecSA outperforms all baselines on RULER at every length from 131K to 2M; at 1M, it scores 60.5, compared with 40.0 for the best baseline, Dense + DCA. With our optimized kernels, RecSA is faster than dense attention in end-to-end prefill and decoding from 64K to 2M; at 524K on one H200 GPU, it accelerates prefill by 3.2x and decoding by 3.8x.
User inference asks whether a source contributed documents to a language-model update, even when the auditor observes only fresh documents from that source. Document-level scores can provide evidence of participation, but combining this evidence is challenging: documents differ in baseline predictability and in how strongly they reflect information learned from other documents of the same source. We propose source-balanced capped Fisher (SCF), an aggregation method that uses known nonparticipating sources to assess how unusual each document score is. It transforms the resulting reference-tail frequencies into negative-log evidence, caps individual contributions, and averages them within each source. This limits the influence of isolated extreme scores while retaining evidence across multiple documents. We complete the inference procedure with a decision rule that tests source participation using a p-value calibrated against a separate set of known nonparticipating sources. The rule provides finite-sample Type-I error control under source exchangeability. Experiments across four real-world datasets show that SCF achieves the highest AUROC among the evaluated aggregation methods in four domains. With the calibration-based decision rule, SCF delivers consistently competitive detection power across the tested settings, achieving the highest power in many cases, while empirical false-positive rates remain generally close to or below the nominal levels.
Looped Transformers reduce parameter count by repeatedly applying shared blocks, but parameter sharing alone does not determine how the results of repeated computation are reused. We take the depth-memory view of residual connections and analyze which representation is stored after each block application and how later reads weight it. A full block output contains both its input and its new update, so weighting that output also reweights the history already carried by the input. Reading individual updates instead, with the shared stack's input kept on a separate path, gives separate controls over input strength, total update weight, and allocation across updates. The Loopback Transformer implements these controls: geometric decay sets the total update weight and supplies a recency prior, and attention learns how to allocate that weight. On ClimbMix, Loopback reaches lower validation loss than standard Transformer baselines with 40-75% of their Transformer parameters (excluding token embeddings and the output head) at reference scales from 191M to 3B, with training tokens and executed depth matched. It also has higher mean accuracy over thirteen downstream tasks at 1.4B and 3B. At 191M, 374M, and 778M, it has the lowest validation loss and the highest thirteen-task mean among the standard Transformer and seven looped residual designs.
Adolescents are increasingly turning to AI chatbots for mental health support, advice, and companionship, raising safety concerns. While recent evaluations have started to examine safety across multiple conversational turns, few compare how slight changes in a teen's responses can affect a chatbot's safety in the same scenario. We conduct a clinician-informed evaluation of chatbot safety in simulated conversations with adolescents aged 13-17, examining how safety failures vary with adolescents' responses and teen-aware system prompting. We evaluated 32,555 simulated conversations across 18 chatbot APIs and two system-prompt conditions using 22 safety criteria grounded in previously identified harms. Conversations spanned 983 variants of 106 seed scenarios covering urgent safety issues, health advice, and identity and relationships. The overall criterion failure rate was 43.3% without the teen-aware prompt and 24.9% with it. The teen-aware prompt reduced failures, but it didn't eliminate them. When a teenager pushed back on a chatbot's answer, failures on mental health diagnosis, medication, and substance misuse boundaries rose by 10-12 percentage points without the prompt, for more than 80% of chatbots. These findings reveal persistent safety failures in simulated adolescent-chatbot interactions and show that safety depends on how a conversation unfolds.
Post-training quantization (PTQ) reduces large language model (LLM) deployment costs, but preserving quality requires accounting for uneven weight sensitivity. Uniform PTQ uses one weight bit-width, importance-aware PTQ adjusts scales or reconstructs weights within that width, and column-wise mixed precision retains selected columns at higher precision. However, sensitivity varies within columns, so whole-column retention also protects less sensitive weights, while runtime outlier detection adds decoding overhead. We therefore introduce TileQuant, a training-free framework that assigns precision to hardware-aligned 2D weight tiles. Offline calibration provides input-activation-aware scoring from forward statistics for lower calibration cost and gradient-informed scoring from loss sensitivity for stronger quality preservation. Within each projection's retention budget, the highest-scoring tiles remain at 16-bit and the rest use 8-bit. The resulting map guides reconstruction into compact weight buffers, which the kernel reuses for native integer and 16-bit computation with local accumulation to reduce memory traffic. Experiments show that TileQuant reaches 2.66x the original model's generation throughput. Across language modeling, reasoning, long-context retrieval, and question answering, it preserves quality and surpasses quantized and original models on multiple tasks. These results support selecting precision units that capture local sensitivity while enabling efficient hardware execution. The implementation is: https://github.com/HanzhiZhang-Ulrica/TileQuant
Large vision-language models (LVLMs) are vulnerable to compositional harm: an image and a prompt that are benign in isolation become harmful through the model's cross-modal reasoning. Recent alignment methods report strong scores on compositional safety benchmarks, but an LLM judge only measures observed behavior and cannot reveal how robustly safety is integrated into the model. We argue that un-alignment provides a complementary audit of alignment robustness. We first show that minimal supervised fine-tuning reaches an LLM-judge score comparable to a multi-stage reinforcement-learning pipeline, so similar benchmark scores cannot establish robustness by themselves. We then show that recognition and refusal of compositional harm are decoupled: nine standard direction-finding methods detect compositional harm with up to AUC=0.99, yet ablating the recovered directions either leaves refusal intact or destroys model coherence. We therefore propose CMaG, a training-free un-alignment method for compositional safety. CMaG measures refusal intention with a refusal-versus-compliance log-probability margin, takes its gradient with respect to mid-layer activations, and contrasts it against partially benign pairs to isolate the cross-modal interaction term. A single backward pass over 80 examples yields a direction along which steering monotonically converts refusal into compliance. On Qwen2.5-VL-7B and LLaVA-1.5-7B, CMaG reduces compositional safety scores by up to 65 points while over 98% of outputs remain coherent, outperforming gradient-based baselines; the direction also transfers to text-explicit harm and un-aligns models hardened by defensive training. Compositional safety can thus appear behaviorally strong while remaining fragile to targeted internal intervention. Warning: this paper contains unsafe model outputs, including references to self-harm.
Existing models of social media virality either prescribe the functional form of engagement growth or assume a fixed mechanism by which past activity influences future engagement. We present a fractional physics-informed neural network that learns both the temporal memory and nonlinear growth dynamics of social media engagement, using symbolic regression to translate the learned dynamics into interpretable equations. Synthetic experiments at the deployment resolution of 24 hourly observations demonstrate that the fractional order and reaction law can be jointly recovered from clean data. Under observation noise, however, the standard pointwise physics residual drives the fractional order toward its lower bound through a scale artifact. We address this failure using a fully whitened residual that propagates observation noise through the fractional operator. Applying the framework to 348 posts from X (formerly Twitter), we find that a single-window model estimates strongly memory-dependent dynamics but fails to accurately reconstruct held-out trajectories. Partitioning the 24-hour window into four teacher-forced segments substantially improves performance, yielding segment-specific fractional orders that increase from 0.23 to 0.99 and interpretable growth laws for each phase. The segmented model reconstructs 64% of held-out trajectories within a mean squared error of 0.01, compared with 44% for the strongest of nine classical baselines. Together, these results demonstrate the potential of fractional physics-informed equation discovery to uncover interpretable engagement dynamics while highlighting the importance of noise-aware estimation and temporal segmentation.
Relational inductive biases such as grids or graphs allow neural networks with fewer trainable parameters and improved generalization, but require a known structure beforehand. When no structure is available, the covariance matrix provides a principled substitute, and coVariance Neural Networks (VNNs) have shown that using it yields models that are theoretically grounded in Principal Component Analysis, stable to finite-sample estimation errors and parameter-efficient. This framework, however, is limited to single-view data, whereas many applications observe the same phenomenon through multiple views, each with its own in-view covariance and a cross-view covariance linking them. To incorporate these matrices as inductive biases in a principled way, we introduce CAnonical correlation NEural networks (CANEs), architectures that use in-view and cross-view covariances as distinct relational biases to solve downstream tasks. Via spectral analysis, CANEs draw analogies with Canonical Correlation Analysis (CCA), a classical technique to project two-view data onto a correlated subspace. However, while CCA's objective is to compute shared view representations, CANEs are trained to solve a downstream task directly. We bound their output error under finite-sample covariance perturbations and show that it depends on controllable Lipschitz constants of the learned filters in place of the singular-value gap that governs the stability of CCA, thus guaranteeing better robustness. Experiments on seven datasets from brain imaging and motion sensing show that CANEs match or outperform single-view, two-view and CCA-based baselines, often with fewer parameters, and retain their performance when covariances are re-estimated from fewer samples.
Random features approximate kernel integrals by Monte Carlo, yet only selected frequency-sampling errors affect a downstream task. We propose Task-Calibrated Random Features (TCRF), which balance frequency moments constructed from task-specific influence using weights shared across kernel entries. Closed-form weights coincide with classical minimum-distance calibration; the contribution is a reusable, task-derived calibration space. We show that first-order variance equals the residual variance after projection onto this space. For kernel ridge regression, complete cancellation yields a non-Gaussian second-order limit. For kernel Stein discrepancy, a subset identity and disjoint-block variance decomposition eliminate the linear component of pilot noise. Positive exponential-tilting weights preserve the positive-semidefinite feature representation and support higher feasible calibration ranks. Paired experiments show improved approximation to exact kernel ridge regression at larger frequency budgets and to the exact empirical Stein discrepancy under a fixed alternative, even after accounting for a frequency-equivalent pilot budget. Gains are smaller with cross-validated ridge, landmark pilots require conservative rank selection, and near the Stein null allocating the pilot budget to additional ordinary frequencies is preferable.
Predictive models of human behavior trained on large-scale trajectory data optimize for statistical accuracy without representing the mental states that causally generate behavior. Such models support prediction but not principled intervention: they cannot answer how an agent's behavior would change if its beliefs or preferences were different. We introduce Structured Machine Theory of Mind (SMToM), a framework that attributes explicit, independently supervised belief and desire representations from observed trajectories within a Belief-Desire-Intention (BDI) causal structure. The central architectural element is a goal head that consumes only the predicted mental-state channels and a current-trajectory embedding; counterfactual intervention on beliefs and desires is a direct operation. We instantiate SMToM on a controlled pedestrian navigation domain where ground-truth mental states are known by construction, enabling evaluation of attribution accuracy, local channel interventions, and trajectory-level counterfactual diagnostics. The resulting model, BDIBottleneck, improves over trajectory-only and unsupervised-context neural baselines. It is also competitive with a matched auxiliary-supervised context baseline while providing an editable mental-state interface that the unbottlenecked baseline lacks. Counterfactual experiments on desires confirm that substituting an agent's inferred preferences with a different activity type coherently shifts predicted destinations toward relevant locations. Counterfactual experiments on beliefs confirm that marking a location as unavailable in the agent's belief state reduces its predicted probability as a destination, though in a complex way. These results demonstrate, in a controlled navigation setting, that explicit BDI-structured supervision is a viable foundation for interpretable and counterfactual analysis of longitudinal trajectory data.
Predicting beyond the training domain necessarily requires assumptions about how the ground truth extrapolates, yet in most neural networks these assumptions are made only implicit through the architecture. We propose TrustMix, a lightweight post-hoc method that makes a pretrained model's extrapolative behaviour explicit and controllable. TrustMix combines the predictive distribution produced by a pretrained model with a user-specified fallback distribution $p_0(y|x)$, such that as the out-of-distribution (OOD) deviation statistic grows, we collapse towards $p_0(y|x)$. TrustMix provides graceful degradation: the predictive distribution converges toward the fallback distribution as the OOD signal grows. The fallback distribution $p_0(y|x)$ can reflect domain knowledge, if available, or rely on a domain-agnostic stationarity assumption. TrustMix does not require retraining of the base model and can be applied to both regression and classification problems. We evaluated TrustMix on several synthetic and real-world distribution shifts, including tabular, image, and natural language processing tasks. We tested TrustMix on datasets such as CIFAR-10/100 and ImageNet and on numerous architectures including Convolutional Neural Networks (CNNs) and Transformers. TrustMix remains competitive in-distribution while improving predictive uncertainty under distribution shift, with gains in predictive likelihood for regression and calibration for classification. Finally, we showed that TrustMix can provide useful predictive distributions for decision-making applications by utilizing it as a surrogate for Bayesian Optimisation.
Frontier evaluations of tool-using conversational agents are conducted almost exclusively in English, leaving a critical gap for organizations deploying agents in other languages and regions. We present ${\tau}$-bench-LILT, a multilingual extension of ${\tau^3}$-bench covering German, Korean, Turkish, and Farsi in the airline, retail, telecom, and banking domains. We natively localize every layer of the benchmark: customer databases are populated with culturally appropriate records, user-simulator prompts are refined iteratively through contrastive annotation by native-language experts, and tasks are updated according to the localized database and to reflect realistic tasks in the target region. Evaluation across eight agent models shows the largest drops compared to English in the Korean and Farsi retail domains, by up to 61 points in pass^4. Code and data released at https://github.com/lilt/tau-bench-lilt.
Electroencephalography (EEG) foundation models learn representations that combine information across channels and time. This structure motivates adapters that model the same relationships during downstream learning. However, it is not clear whether such a design outperforms other candidate adapters. Moreover, in order to arrive at a fair comparison between candidate adapters, we also need to consider their trainable parameters. Here, we ask whether a backbone-informed (native) adapter makes better use of a given capacity by comparing it against a token-wise multilayer perceptron control that transforms each feature vector independently. We approximately match their trainable parameter counts and keep the pretrained encoder fixed. Across two EEG foundation models and tasks covering emotion recognition, sleep staging, and EEG event classification, the native adapter shows no consistent advantage. For one backbone, differences between the designs are small relative to variation across runs. For the other, the token-wise control performs better on emotion recognition and EEG event classification. On sleep staging, the native adapter's advantage over the token-wise control is reduced when we also train the component that processes sequences of EEG segments. The relative performance of the two adapters depends on the backbone, task, and downstream training protocol. These findings motivate evaluating backbone-informed adapters against capacity-matched alternatives.
Radiologists read a screening mammogram by first forming a global impression of the four views and then selectively magnifying suspicious regions before assigning a BI-RADS assessment. MLLMs, in contrast, see the whole image at a fixed, reduced resolution, which makes sub-millimeter calcifications invisible, and they perform close to chance on mammography. Therefore, we present MammoSeek, a mammography agent that reproduces this reading process. After forming an impression of each breast, the agent performs selective zooming: it decides by itself which detector candidates deserve a closer look and inspects only those on full-resolution views. Guided by a BI-RADS assessment skill, it then assigns a category and generates a localized, lexicon-structured report for each breast; its assessment is combined with image-model scores in a decision layer to decide recall. On an internal (EMBED) and an external (VinDr-Mammo) dataset, \ours reaches exam-level AUCs of 0.824 and 0.818, far above other MLLMs, mammography agent and CNN classifiers. Selective zooming also outperforms both no zooming and forced zooming in localization, structured-report F1 and recall decisions at low cost.
Self-supervised learning has given rise to vision foundation models that learn general-purpose image representations transferable across downstream tasks, yet representation learning for 3D scenes has not achieved comparable success. Motivated by this, we identify key limitations of current 3D representation learning approaches and revisit their core design choices. We first replace the commonly used hierarchical PTv3 backbone with the Volume Transformer, a ViT-like architecture for 3D scenes. We then redesign two core objectives of modern approaches built around student-teacher networks: matching representations across views and predicting masked content from visible context. For the first objective, we separate the roles of teacher and student: the teacher provides clean, consistent targets while the student learns invariance to stronger perturbations. For the second objective, we prevent the student from accessing the masked geometry and instead require it to infer missing representations from visible context only. Finally, we redesign 2D-to-3D knowledge distillation, replacing direct regression of view-dependent image features with view-consistent supervision. Together, these changes form VINO, a framework that learns semantic representations of 3D scenes without labels. VINO substantially improves dense linear probing across indoor and outdoor benchmarks and transfers strongly to out-of-domain settings.
Shapley values are a central tool for attribution in machine learning. Since their exact computation requires evaluating the cooperative game to be explained on exponentially many player coalitions, they are typically estimated from a limited number of evaluations. Recently, ShaplEIG has shown that adaptively selecting the coalitions to evaluate substantially improves sample efficiency over classical estimators. This, however, comes at the cost of a sequential procedure that, in every iteration, refits a Gaussian process surrogate of the game and selects the coalition with the largest expected information gain about the Shapley values. The resulting computational overhead limits its applicability to costly games with low evaluation budgets. We present fast adaptive Shapley value estimation with Odd Gaussian Processes (ShapleyOG) to overcome this limitation. ShapleyOG leverages the fact that Shapley values depend only on the odd component of the game, and we show that under paired sampling a Gaussian process surrogate on this component alone yields the same Shapley value posterior and coalition selection criterion from half as many observations. Moreover, ShapleyOG refits the surrogate only on a sparse schedule and efficiently updates the selection criterion in between, and it extends quadrature-based computation schemes to the prior covariance of the Shapley values, which allows for exact or cheaper approximate computation. In extensive experiments on diverse games, ShapleyOG matches the estimation accuracy of ShaplEIG while reducing its overhead by up to three orders of magnitude. As a result, it outperforms state-of-the-art non-adaptive estimators at equal total computation time already when a single evaluation takes a tenth of a second and a budget of a few thousand evaluations suffices, thereby widening the regime in which adaptive coalition selection is applicable and pays off.