Respiratory diseases remain a leading cause of global mortality, where timely and accurate diagnosis is critical to improving patient outcomes and reducing healthcare burdens. While prior work has explored audio-based models for respiratory disease detection, such unimodal approaches often suffer from limited generalizability and diagnostic precision. In this paper, we propose RespiraMFM, a Multimodal Foundation Model that integrates respiratory sounds with patient medical history and symptoms to enhance diagnostic accuracy and disease detection capabilities. We introduce an effective contrastive alignment strategy for audio-text multimodal integration, allowing the model to learn better cross-modal representations between respiratory sounds and corresponding textual clinical information. We evaluate RespiraMFM across five major respiratory diseases using seven real-world datasets in both supervised fine-tuning and zero-shot settings, achieving a 9.15% improvement in AUROC on supervised tasks and a 20.98% gain on zero-shot tasks over existing baselines. These findings underscore the potential of our framework to advance early diagnosis and improve clinical decision-making in respiratory disease management.

Continuous diffusion for categorical data is a framework belonging to the diffusion family and aiming at generating discrete data. The scientific interest to such models has been constantly increasing these days because researchers try to achieve a challenging goal of finding reasonable alternatives to autoregressive large language models. In this paper, we study the properties of the structure of the latent space corresponding to discrete tokens expressed in terms of Kullback-Leibler divergence on diffusion path measures and accuracy of the correct token prediction by the optimally trained diffusion model. We find that FSQ tokenization scheme has the latent space structure with the properties that make it best suited for continuous diffusion for categorical data as verified through rigorous theoretical analysis and numerical experiments. To validate our findings in real-life scenario, we train several text-to-speech diffusion models having speech tokens as intermediate acoustic features, and show that the one based on FSQ tokens indeed performs the best, and, moreover, it outperforms its strong LLM-based counterpart, at the same time being significantly smaller and faster.

Training large language models (LLMs) as autonomous agents via reinforcement learning (RL) has enabled frontier models to achieve superhuman performance in long-horizon tasks. However, existing RL algorithms operate at the trajectory level, performing policy optimization only after collecting complete episode rollouts. This coarse-grained approach faces fundamental challenges in multi-turn agent settings where rewards are sparse, delayed, and credit assignment across individual steps is critical. In this work, we propose \textbf{State-Score-Supervised Policy Optimization (3SPO)}, a novel RL algorithm that performs post-step policy optimization with dynamic state score supervision. At each step, 3SPO computes the state score based on historical success rates, supervising step-wise credit assignment, adaptive rollout and post-step policy optimization without requiring value function estimation or additional auxiliary models. Theoretically, under a per-state bandit abstraction, we show that the proposed score-supervised allocation mechanism achieves logarithmic allocation regret and provide sample-complexity guarantees for action identification, score distinguishability, and filtering stability. Experiments on ALFWorld and WebShop with Qwen2.5-1.5B/7B-Instruct show that 3SPO consistently outperforms GRPO by $+22.6\%$ on ALFWorld and $+15.6$ points on WebShop, while using comparable resources to achieve $2.4\times$ more state exploration and $1.8\times$ faster convergence. Code is available at https://github.com/genalyu/3SPO.

We present HydraCIL, a decoupled continual learning model based on prototype-guided multi-head classifiers, targeting sustainable deployment in embedded and resource-constrained environments. While most Class-Incremental Learning (CIL) methods rely on powerful hardware and long retraining cycles, real-world systems, such as robots or edge AI devices, must adapt quickly with limited resources. HydraCIL addresses this gap by freezing the backbone and decoupling feature extraction from learning. For each task, features are extracted once and a lightweight, task-specific classifier head is created, avoiding costly backbone retraining. At inference, HydraCIL selects the appropriate head via similarity with prototypes. Experiments on CIFAR-100, ImageNet-100, CoRe50, and Flowers102 datasets show that HydraCIL matches or outperforms state-of-the-art CIL methods while significantly reducing training time and carbon footprint, making it a practical solution for continual learning in real-world and embedded settings, where energy efficiency and rapid adaptation are critical.

Precipitation nowcasting is increasingly being approached with deep learning models that learn directly from recent radar observations. Although such models can efficiently capture short-term precipitation motion, they often lack broader contextual information about the meteorological conditions under which rainfall develops. This paper investigates whether lightweight temporal context can improve radar-based nowcasting, particularly for high-intensity rainfall. We propose the Time-Aware Small-Attention U-Net (TA-SmaAt-UNet), which extends the core SmaAt-UNet model with temporal conditioning layers that use cyclical encodings of time-of-day and time-of-year to modulate intermediate feature representations. Experiments on KNMI radar precipitation data show that temporal conditioning is most beneficial for rare, high-intensity precipitation events, while also improving the representation of seasonal variability and predicted rainfall-intensity distributions. A layer conductance analysis further indicates that the added temporal conditioning layers are actively used by the model despite their small parameter cost. These findings suggest that simple, physically motivated temporal context can improve the realism and reliability of deep learning-based precipitation nowcasts. The implementation of our models and training setup is available on \href{https://github.com/gijsvn/TA-SmaAt-UNet}{GitHub}.

In mixed-traffic environments where autonomous and human-driven vehicles may co-exist, motion planning for autonomous vehicles requires anticipating the future behaviors of surrounding human drivers. Existing reinforcement learning-based methods generally directly incorporate the predicted human intents into the observation to enable a proactive planning. However, human intent is inherently uncertain due to the behavioral diversity, perception noise, and partial observability. Treating predicted intends as deterministic states can result in unsafe decisions for autonomous vehicles. To address this problem, we propose Uncertainty-Aware Motion Planning (UAMP), which incorporates uncertainty in human intent prediction for AV decision-making. Specifically, UAMP first introduces a proximity-aware uncertainty estimator to quantify the interaction-conditioned intent uncertainty and constructs an uncertainty-guided joint intent distribution over surrounding human-driven vehicles. Within this uncertainty set, UAMP further introduces Uncertainty-Calibrated Value Learning (UCVL) to correct value function learning biases arising from directly incorporating uncertain human intent predictions into the observation. Extensive experiments in various mixed-traffic scenarios show that UAMP significantly improves safety and driving comfort, while maintaining traffic efficiency compared with existing approaches. The code is released at https://anonymous.4open.science/r/UAMP-5638.

Large models for time-series forecasting have been emerged as a promising paradigm for training models on heterogeneous collections of signals. These models typically rely on causal autoregressive architectures, where each observation is sequentially predicted from past. In practice, real-world time-series exhibit non-stationarities, which significantly influence predictive performance. To mitigate this, normalization is commonly employed. However, in efficient causal settings it might induce information leakage from future observations during training. Recent alternatives, including causal normalization and statistics computed from initial observations, have been proposed to address this issue, but their practical implications remain insufficiently understood. In this work, we evaluate normalization strategies for transformer-based large time-series models trained with patching and efficient causal strategy. We showcase that normalization choice significantly influences both training convergence and forecasting performance.

During the training of large Transformer models, attention masks regulate the scope and direction of information flow across a sequence. Numerous mask variants exist, and operators such as FlexAttention already support arbitrary attention masks. Nevertheless, a systematic formal analysis of the information-flow structure induced by arbitrary masks has been missing. This paper develops a complete theoretical framework. We prove that, with sufficient depth, the information flow of a multi-layer Transformer converges to a Hasse diagram -- a directed acyclic graph representing a partial order. Building on this, we recast the design of parallel training tasks as the problem of finding a minimal common supergraph of Hasse diagrams, and we establish a criterion for the minimal common supergraph. This yields a constructive method to derive attention masks directly from a family of tasks. Applying the framework, we design two novel masks: a block-generation attention mask that ensures training-inference consistency (Block Two-Stream Attention), and a fully supervised bidirectional attention mask (Butterfly Attention). These results demonstrate the framework's capacity to discover new structures.

Data-driven PDE surrogates are trained with data produced by numerical PDE solvers. However, when the surrogate's goal is to generalize across a wide range of PDE configurations (e.g., initial conditions and physical coefficients), generating a representative training set is non-trivial. Uniform sampling of configuration parameters often under-represents trajectories exhibiting challenging dynamics, leading to high prediction errors and large error variance in the trained surrogate. Online training, where data generation and surrogate training are coupled, offers a natural advantage by allowing solver parameters to be steered on-the-fly. To efficiently exploit this capability, we introduce Online Generative Active Sampling (OGAS), an active learning method that reactively learns the relationship between configuration parameters and surrogate performance to control the sampling distribution. OGAS trains a fast diffusion model in parallel to the surrogate to act as a conditional sampler, mapping a surrogate-derived difficulty signal (e.g., loss or uncertainty) to configuration parameters. By actively drawing target signals from a prior biased toward high difficulty, OGAS continuously steers data generation toward challenging regimes without delaying the training workflow. We evaluate OGAS across 2D PDEs with distinct challenging dynamics (Kuramoto-Sivashinsky, Navier-Stokes, Gray-Scott) and up to 308 parameters, using multiple surrogate architectures. Across all settings, OGAS consistently improves tail statistics, yielding substantial reductions in errors above the 99th percentile and overall error dispersion compared to uniform sampling. While prioritizing challenging trajectories introduces a trade-off with average error, OGAS effectively ensures worst-case reliability of trained surrogates with negligible wall-time overhead.

Dictionary learning methods like Sparse Autoencoders (SAEs) and crosscoders attempt to explain a model by decomposing its activations into independent features. Interactions between features hence induce errors in the reconstruction. We formalize this intuition via compact proofs and make five contributions. First, we show how, \textit{in principle}, a compact proof of model performance can be constructed using a crosscoder. Second, we show that an error term arising in this proof can naturally be interpreted as a measure of interaction between crosscoder features and provide an explicit expression for the interaction term in the Multi-Layer Perceptron (MLP) layers. We then provide three applications of this new interaction measure. In our third contribution we show that the interaction term itself can be used as a differentiable loss penalty. Applying this penalty, we can achieve ``computationally sparse'' crosscoders that retain $60\%$ of MLP performance when only keeping a single feature at each datapoint and neuron, compared to $10\%$ in standard crosscoders. We then show that clustering according to our interaction measure provides semantically meaningful feature clusters, and finally that sleeper agents have significant interactions. Code is available at https://github.com/chainik1125/crosscoders-feature-interactions/tree/arxiv.

We introduce RKSC (Reasoning-Aware KV Cache Sharing), a training-free inference framework that eliminates two structural redundancies in multi-branch LLM reasoning pipelines. ASKS (Attention-Similarity KV Sharing) computes the prefix KV cache once and broadcasts it to all semantically similar branches via hidden-state cosine similarity, strictly generalising the token-exact prefix caching used by vLLM and SGLang. CGEE (Confidence-Gated Early Exit) applies two complementary exit mechanisms: (1) it skips the verification forward pass entirely when generation confidence is decisive across branches, and (2) it terminates the verification pass at an intermediate layer when per-layer entropy stabilises, using lightweight hooks on the transformer backbone. RSBCM (Reasoning-Selective Block Cache Manager) prevents unbounded cache growth via attention-weighted depth-priority eviction. Across five model families (7B-10B), four benchmarks, and 1,000 evaluated problems, RKSC achieves a mean speedup of 3.008x over the No-KV baseline (peak 3.990x), a 1.66x mean improvement over vLLM-equivalent prefix caching, with a CGEE-induced error rate of only 0.37% (6 errors out of 1,616 verify calls). No fine-tuning or architecture changes are required. Code is available at https://github.com/AnirudhSekar/RKSC.

World models are now built on substantially different computational substrates. Latent recurrent state-space models such as PlaNet and the Dreamer family compress observations into recurrent states; token-based models such as IRIS quantize observations into a learned codebook and predict autoregressively with a transformer; and joint-embedding predictive architectures such as I-JEPA predict in a learned latent space with no pixel decoder. The interpretability methods applied to these models, including probing, activation patching, sparse autoencoders, and surprise analysis, share a common set of primitives, yet they are re-implemented from scratch for each architecture because existing hook-and-cache tooling assumes a transformer language model with no notion of actions, environment steps, or imagined rollouts. We argue that this fragmentation reflects the tooling rather than the models, and that the shared structure of world models is captured by a small typed interface. We present WorldModelLens, an open-source interpretability substrate organized around a capability-typed adapter: every model implements four required methods (encode, transition, initial state, sample) and declares a set of optional heads (decode, reward, continue, actor, critic) through an explicit capability descriptor, so that reinforcement-learning and self-supervised world models are first-class without either imitating the other. A single hook and cache layer exposes time-indexed activations, imagination rollouts, and intervention replay over this interface, allowing each analysis to be written once.

Feature selection is critical for network intrusion detection systems (NIDS) operating under high-dimensional, highly imbalanced traffic, as found in operational and defense networks. Traditional filter methods rank features using global statistics computed symmetrically across classes and thus fail to capture the asymmetry of intrusion detection, where attacks are best characterized as deviations from dominant benign traffic. We propose benign-anchored Classwise Mean Deviation (nCMD), a lightweight and interpretable method that scores feature relevance based on the deviation of attack-class distributions from the benign-class mean, rather than a globally biased reference. This approach aligns feature selection with the operational semantics of NIDS at no additional computational cost. Across four benchmark datasets (CICIDS2017, CICDDoS2019, NSL-KDD, and UNSW-NB15), multiple feature budgets, and three downstream classifiers, nCMD matches or exceeds classical filter baselines in macro-averaged F1-score. It achieves the best result on three of the four datasets and under every classifier, with the strongest improvements observed under tight feature budgets and severe class imbalance. These results support benign-anchored ranking as a scalable and interpretable preprocessing component for resource-constrained NIDS.

Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has become a standard pipeline for Large Language Model (LLM) post-training. SFT is expected to provide a useful behavioral prior for RL to further enhance model capabilities. However, checkpoints with excessive SFT often show limited improvement during RL. We attribute this failure to the loss of model plasticity: the reduced ability of an SFT-initialized policy to be effectively reshaped by subsequent RL. To better understand this phenomenon, we conduct detailed analysis from multiple perspectives, including parameter changes, output spaces, and RL optimization dynamics. Our results show that models from excessive SFT tend to produce over-confident token distributions and exhibit sharp parameter landscapes, which make them harder to optimize in the RL stage. To enable a more robust SFT-to-RL handoff, we propose \texttt{Rejuvenation}, a simple yet effective method that restores plasticity while preserving useful SFT-acquired priors. Rejuvenation leverages base-anchored model fusion to reduce excessive SFT-induced drift with targeted neuron reset to mitigate model rigidity. Experimental results on both math reasoning tasks and agentic tasks demonstrate that our approach consistently improves RL performance on over-trained SFT models, while also enhancing generalization to out-of-distribution tasks.

Physical reservoir computing harnesses nonlinear mechanical dynamics but, by convention, freezes the substrate and trains only a linear readout, presuming the substrate is not usefully trainable. We revisit that premise for networks of nonlinear oscillators whose mass, damping, and stiffness are learned end-to-end through a symplectic integrator. Our central result is a trilemma: memory horizon, gradient stability, and dynamical expressivity cannot be simultaneously maximized, because all three are governed by the damping. The backward gradient decays at a rate set by the damping, capping how far back credit can propagate, while forward sensitivities grow exponentially in the largest Lyapunov exponent, so usable gradients require damping above a stability floor. Since the Lyapunov exponent falls as damping rises while the memory ceiling falls as the horizon grows, stable training is confined to a band that contracts with horizon and closes at a critical point. We test every step on a twenty-oscillator network. A damping sweep finds the largest Lyapunov exponent monotone and crossing zero at a well-defined stability floor, confirming the theorem's key assumption. A compute-matched comparison of learned versus frozen substrate on delayed recall across nine horizons shows the learned substrate dominating at short horizons and the advantage closing and reversing near a horizon of eleven steps, the predicted signature of band closure; trained models settle near the stability floor, seeking the edge of chaos unprompted. The analytic ceiling overestimates the empirical crossover roughly fivefold, a gap between detectable and learnable gradient that we report rather than tune away. The contribution is a confirmed account of when training a physical substrate beats freezing it.

The Forward-Forward (FF) algorithm offers a biologically inspired alternative to backpropagation by replacing gradient-based credit assignment with local, forward-only objectives. While recent extensions have adapted FF to convolutional neural networks (CNNs), existing formulations rely on static channel-class partitions and struggle to perform effectively in complex tasks. In this work, we introduce a learnable channel-class assignment mechanism that enables adaptive, data-driven specialization of convolutional channels, supported by entropy and orthogonality regularization to promote learning performance. We further propose a loss-aware layer contribution strategy that adaptively weights intermediate-layer predictions based on their validation performance, enhancing the effectiveness of forward-only inference. Integrated into residual CNNs, the proposed method achieves consistently superior performance across CIFAR-10, CIFAR-100, and Tiny-ImageNet compared to existing similar forward-only methods. Notably, it establishes new state-of-the-art performance among FF-based models, substantially narrowing the gap with backpropagation. These findings demonstrate that introducing learnable channel specialization and layer contribution weighting significantly enhances the representational capacity of forward-only learning in deep CNNs.

Post-training quantization (PTQ) is one of the most practical ways to reduce the serving cost of Large Language Models (LLMs), but activation quantization remains difficult because outlier-dominated channels lead to large quantization errors. This paper investigates whether part of this degradation is caused by over-migration in scaling-based equivalent transformations. We introduce a quantile-robust scaling policy for SmoothRot-style transforms by replacing max-based activation statistics with high quantiles, and we complement it with constrained gradient-based optimization of channel scales. On LLaMA-3.2-1B under W4A4 quantization, quantile-only policy search improves selected-layer error by 11.1% over the SmoothRot baseline, joint (alpha, q) search improves it by 12%, and training reaches 18.5%. Replaying the best selected-layer policy on all decoder-block down-projection layers reduces the corresponding full-layer mean error from 97.51 to 78.08 (19.9%). The results show that robust migration control and lightweight scale learning provide consistent gains over max-based fixed policies while preserving the equivalent-transform framework.

Sampling from the sequence-level power distribution $p^α$ elicits RL-level reasoning from base language models without any parameter updates, but the standard Metropolis--Hastings (MH), a Markov Chain Monte Carlo (MCMC) sampler, is both expensive and slow-mixing. We trace both to a structural mismatch: $p^α$ mainly departs from $p$ at a sparse, spatially clustered set of high-entropy decision points, yet MH proposes resampling positions uniformly along the prefix -- wasting compute on near-degenerate conditionals while under-mixing precisely where modes diverge. We propose Entropy-Guided Power Sampling (EGPS), a training-free and verifier-free sampler that re-derives its proposal from token-level entropy already in the forward pass. EGPS skips deterministic blocks, localizes each MCMC move to a high-entropy neighborhood, and applies Multiple-Try Metropolis at decision points -- making sampling cost scale with \emph{entropy mass rather than sequence length}. On Qwen2.5-Math-7B, EGPS reaches best or tied-best accuracy on all three benchmarks (MATH500 $75.8\%$, HumanEval $62.2\%$, GPQA $42.4\%$) at up to a $12.6\times$ wall-clock speedup over the MH baseline.

Large audio-language models (LALMs) increasingly use explicit reasoning traces for complex audio understanding, yet the evaluation of reasoning quality remains underexplored. Although process-level benchmarks for process reward models (PRMs) have advanced reasoning evaluation in text and multi-modal domains, comparable evaluation for audio reasoning remains limited. In this paper, we present AudioProcessBench, a comprehensive benchmark for step-level process error identification in audio reasoning. AudioProcessBench contains diverse reasoning traces generated by 6 audio and omni language models. Each trace is segmented into discrete reasoning steps and annotated with binary step correctness and fine-grained error types. Our benchmark evaluates models under three complementary paradigms: (1) step correctness identification, (2) error-type-conditioned detection for diagnosing audio-specific verifier capacities, and (3) chain-level aggregation, where verifiers select or aggregate among multiple reasoning traces for the same question. This design enables a systematic analysis of whether current models can detect process errors, whether their weaknesses differ across audio-specific error types, and whether process verification translates into improved answer selection. AudioProcessBench provides a testbed for future research on audio reasoning verifiers, process reward models, and reliable omni-modal reasoning.

Deploying deep neural networks on memory-constrained edge accelerators is bottlenecked by per-inference off-chip weight transfer rather than computation: the dense network cannot be retained on-chip, and every parameter must be loaded for every input. Existing model compression reduces this transfer only at the cost of permanent capacity loss. We propose Sigma-Branch (SigmaB), a framework that restructures a pretrained dense network into a hierarchical binary tree composed of a shared backbone, hierarchical routers, and specialized leaves. Pretrained weights are distributed across the tree via activation-based spherical k-means clustering, which jointly initializes router weights and per-branch channel allocations; soft-routing fine-tuning then aligns each leaf with its routed input subset. At inference, the resulting network executes only a single root-to-leaf path, reducing the active-parameter footprint while storing the complete dense parameter set in memory. Across CIFAR-100 / ResNet-50, ImageNet-1K / ResNet-50, and ModelNet40 / PointNet++, SigmaB-Net reduces per-inference active parameters by 58-60% while remaining within 1.72 percentage points (pp) of the dense baseline Top-1. At comparable ImageNet-1K Top-1, the active-parameter reduction exceeds static structured pruning (FPGM, HRank) by 14-23 pp. The cross-modal evaluation, spanning 2D vision and 3D point-cloud backbones, substantiates a framework-level claim that decouples per-inference memory traffic from the total parameter count.

Neural operators such as the Fourier Neural Operator (FNO) have emerged as powerful surrogates for solving partial differential equations (PDEs), achieving speedups of several orders of magnitude over traditional numerical solvers. However, deploying these models in safety-critical engineering applications -- such as thermal management of electronic components and battery systems -- requires not only accurate point predictions but also rigorous uncertainty guarantees. Existing uncertainty quantification (UQ) methods for neural operators, including Monte Carlo Dropout and Deep Ensembles, provide only relative uncertainty estimates without formal coverage guarantees. In this work, we propose the first application of split conformal prediction to neural operator-based physics simulation, providing distribution-free prediction intervals with finite-sample coverage guarantees. We further introduce a normalized conformal prediction scheme that leverages MC Dropout uncertainty to produce adaptive-width intervals, yielding tighter intervals in regions of low uncertainty and wider intervals where the model is less certain. Full-scale experiments (33.7M parameters, 800 training samples, 5 ensemble members, NVIDIA V100) on steady-state heat conduction benchmarks demonstrate that our method achieves 89.1% empirical coverage at the target level of alpha=0.1, while producing spatially adaptive prediction intervals that reflect the underlying physical uncertainty structure. We also provide an uncertainty decomposition framework that separates epistemic uncertainty (68% of total) from aleatoric uncertainty (32% of total), offering actionable guidance for data collection and model improvement. Our method is implemented in an open-source platform with REST API endpoints and interactive 3D visualization.

Perceptual uncertainty is a central challenge for heterogeneous robot teams operating in unstructured outdoor environments, where no single viewpoint affords reliable scene understanding. Perceptual uncertainty, arising from sources such as occlusions, manifests differently across robot viewpoints depending on scene structure. Detecting and resolving sources of perceptual uncertainty requires both scene-based contextual reasoning and capability-aware robot allocation. While vision-language models provide strong semantic priors for both, they are computationally prohibitive for onboard inference and lack calibrated uncertainty quantification. We introduce Co-GLANCE, a real-time onboard perception and decision-making system for uncertainty resolution in heterogeneous robot teams. Co-GLANCE distills the semantic reasoning capabilities of a vision-language model into an end-to-end model for occlusion segmentation and robot allocation, eliminating the need for cloud-based inference. To quantify perceptual uncertainty, Co-GLANCE combines conformal prediction with selective abstention to provide statistically valid coverage guarantees for segmentation, robot allocation, and detection outputs. These calibrated uncertainty estimates directly trigger active perception, dispatching the most appropriate robot to acquire informative viewpoints and resolve uncertainty. Across real-world scenarios, Co-GLANCE outperforms cloud-based vision-language model baselines in occlusion segmentation and robot allocation accuracy by 25% and 36%, respectively, while reducing per-frame inference latency 350x. We also release an air-ground dataset for future research. Code, videos, and dataset available at https://co-glance.github.io/ .

Multivariate time series forecasting requires capturing the continuously evolving correlation structure among interacting variables. Existing state-space models process time series by scanning tokenized temporal or spatial sequences, discarding the evolutionary geometric structure. We address this limitation by introducing manifold constraints into state-space modeling: treating the cross-variable correlation structure as a continuous trajectory on the symmetric positive definite manifold, whose Riemannian geometric features, tangent space linearity, and Frechet mean centrality act as a principled geometric regularizer that guides and stabilizes the selective scanning dynamics of SSMs. We propose SPDM, a geometry-aware SSM architecture that realizes this principle through two cooperating mechanisms: a manifold trajectory path that projects dynamically evolving covariance matrices from the SPD manifold to a Euclidean tangent space, and a geometric gating scheme that directly modulates SSM's internal selective parameters based on geometric signals derived from the manifold trajectory. The parameterization preserves the linear-time complexity of the Mamba parallel scan while embedding rich structural constraints, making the architecture preserve prediction accuracy and computational efficiency simultaneously. Extensive experiments on eleven real-world benchmark datasets establish state-of-the-art forecasting performance, and further studies confirm that geometrically constrained state-space dynamics are the dominant architectural factor behind its performance gains.

Multi-turn LLM agents fan short queries into long trajectories of tool calls, search results, and intermediate reasoning. Both KV memory and KV read bandwidth grow by orders of magnitude across a single trajectory, making the key-value (KV) cache, not parameter compute, the dominant serving bottleneck for long-horizon agents. We introduce IntentKV, learned KV pruning that keeps the base LLM frozen. IntentKV maintains a session-level QueryMemory of cross-turn intent, scores live history tokens with a memory-attention rule, and adds a zero-initialized residual head with cross-attention over current-query K-vectors. To stay composable with prefix caches, eviction is a slot-map redirection: dropped positions route to a sentinel dead slot while surviving K/V rows, RoPE phases, and slot identities stay in place. IntentKV matches the no-pruning full-cache baseline with almost no accuracy drop under tight KV budgets: at an 8k KV budget, mean peak request tokens drop 23.9% on Qwen3-8B and 30.7% on Qwen2.5-14B. On the 100 longest BCP queries that all methods complete on Qwen2.5-14B, IntentKV-8k further cuts worst-case peak request tokens from 92.3k to 20.5k, a 77.8% reduction, and worst-case raw KV reads from 411M to 31M, a 92.6% reduction.

Choosing the wrong synthetic generator for time-series foundation model pretraining is costly: under identical training budgets, the best and worst generators produce up to a $2\times$ gap in forecasting error, yet the field has no principled way to make this choice. The problem is compounded by the fact that generator rankings are not stable across architectures: across 11 generator families evaluated on Chronos-T5-Mini and Moirai-Small trained from scratch, we find that which generators are useful depends on the model architecture. Rather than solving the generator selection problem, we sidestep it: a simple equal-weight mixture of all generators matches or beats the best individual generator for both architectures, and composing this mixture with real data yields the strongest pretraining corpora overall. Synthetic pretraining is therefore a corpus composition problem, not a generator selection problem, and composition choices should be validated per model family rather than assumed to transfer.

Multimodal clinical learning is increasingly important for integrating diverse patient data, including imaging, text, and personalised health records. However, it faces two fundamental challenges: i) modality missingness, where arbitrary subsets of modalities are unavailable at a given patient visit, ii) longitudinal dynamics, where the diagnostic significance of an observation depends on the patient's evolving disease trajectory over time. Existing methods address these challenges in isolation: missing-modality frameworks treat each visit as an independent static snapshot and discard temporal context, while longitudinal models often assume complete modality availability and degrade under systematic modality incompleteness. We propose LongMoE (Longitudinal Mixture-of-Experts), the unified framework to jointly address both challenges. LongMoE combines a context-aware imputation module with an attentional tokenization module that captures frequency-domain temporal patterns across irregular visit sequences, a trajectory-aware encoder for modeling disease progression, and context-conditioned Sparse MoE routing for patient-specific expert selection. Experiments on ADNI, OASIS-3, and MIMIC-IV show that LongMoE improves robustness under missing or weak contemporaneous modalities and remains competitive in full-modality settings, establishing a strong foundation for longitudinally-aware multimodal clinical learning.

Long-term memory is the missing layer for LLM agents: across sessions they forget, and the common workaround -- replaying the whole history into the prompt -- is expensive, slow, and, as distractors accumulate, less accurate. Most memory systems win on cost or latency but still lose to the full-context baseline on accuracy, and benchmark numbers are reported on inconsistent, non-reproducible harnesses, so one system appears at wildly different scores across sources. We present Engram, an open-source, dual-process memory engine on a bi-temporal data model. A fast write path appends lossless episodes with no LLM on the critical path; an asynchronous path extracts atomic (subject, predicate, object) facts, builds a bi-temporal knowledge graph, and resolves contradictions without an LLM call per fact -- invalidating, never deleting, so every fact keeps provenance and a supersession chain. A hybrid read path fuses dense, lexical, graph, and recency/salience signals, applies a point-in-time ("as-of") filter, and assembles a compact, provenance-tagged context. On the full 500-question LongMemEval_S, graded by the official category-specific judge, Engram's lean configuration -- answering from a ~9.6k-token retrieved slice, never the full history -- scores 83.6% vs. 73.2% for full-context (+10.4 points, McNemar p < 10^-6) at ~8x fewer tokens (9.6k vs. 79k), with 0/500 errored. The gain needs a hybrid read path: facts alone lose recall, while facts plus retrieved chunks recover detail. We also contribute a neutral, in-repo evaluation harness with the official judge baked in and the full-context baseline in every table, publish the raw per-question logs, and document the measurement-integrity pitfalls (truncation, home-grown judges, full-history leaks) that silently distort memory benchmarks. Every number ships with a command to reproduce it.

A central goal of mechanistic interpretability is to identify which internal components causally drive a language model's behavior. Because these importance estimates serve as the evidence for identifying circuits, systematic errors can lead to the misidentification of the underlying mechanisms. While activation patching provides a gold-standard causal metric, its computational cost is prohibitive at scale. Practitioners instead rely on attribution patching, a gradient-based, first-order approximation whose reliability remains poorly understood. In this work, we characterize the source of this unreliability, demonstrating that the dominant error stems from the non-linearities in the downstream network rather than local curvature at the patched component. This insight yields three practical tools: (i) a reliability score to detect untrustworthy estimates, (ii) error bounds quantifying potential attribution mis-specifications, and (iii) a Hessian-vector-product (HVP) correction that eliminates the leading-order error with only one additional backward pass. In evaluations across five model families (124M-9B parameters) and both random-token and naturalistic (name-swap) perturbations, HVP is the only second-order correction feasible at larger scale, where standard baselines like Integrated Gradients become computationally prohibitive. In comparative experiments, a multi-step HVP variant matches or exceeds the accuracy of Integrated Gradients at significantly lower compute, outperforming prior second-order baselines. These improvements lead to higher-fidelity circuit recovery on standard benchmarks and support a Screen-Flag-Fix workflow that targets computational effort only toward the components flagged as unreliable.

Cancer treatment planning requires decisions across multiple clinical dimensions at once. Clinicians must determine whether a patient should receive targeted molecular therapy, radiation therapy, and whether they are likely to survive beyond six months. Existing pathway-informed deep learning models have been developed and tested in isolation, making fair comparison across architectures impossible. We present the first unified benchmark for pathway-guided therapy response modeling, evaluating three biologically informed architectures, BINN, GraphPath, and PATH, across five cancer cohorts drawn from The Cancer Genome Atlas, representing 2,622 patients encoded using Reactome pathway activity scores. Each model is trained jointly on all three clinical outcomes under identical data and evaluation conditions, the first study to treat pathway-structured deep learning as a combined therapy and survival prediction problem. Our results show that no single architecture wins across all tasks: PATH performs best for targeted molecular therapy prediction overall, BINN is most reliable for survival prediction, and no model produces useful predictions for radiation therapy, as the key drivers of that decision are clinical variables not captured in gene expression data. Most strikingly, GraphPath achieves an AUROC of 0.92 on prostate targeted molecular therapy prediction, the highest score in the entire benchmark, demonstrating that lateral co-regulation structure produces exceptional discriminative power when matched to a cohort with a narrow targetable driver programme, even under conditions of extreme class imbalance at only 11\% positive prevalence.

Across contemplative, philosophical, and psychological accounts, human consciousness is often described along a similar spectrum, ranging from reactive and self-focused patterns to more integrative and coherent ones. Understanding whether language models encode such a structured, human-interpretable consciousness spectrum in representation space is important for model guidance, evaluation and alignment. In this work, we study the geometric structure and dynamics of patterns along this spectrum in transformer embedding spaces. We show that embeddings exhibit a globally organized geometry aligned with this spectrum: sentences associated with similar states cluster into locally coherent regions, forming a structured manifold. In particular, higher-level and lower-level regions exhibit convexity-like stability, while intermediate regions form a transition corridor. Dynamically, both utility-guided and geometry-only greedy trajectories consistently traverse from lower- to higher-level regions, passing through intermediate tiers, indicating that navigability is an intrinsic property of the representation space, guided but not dictated by a global directional signal. These results suggest that embedding spaces encode structured and navigable geometry aligned with a hypothesized consciousness-spectrum taxonomy, broadly inspired by recurring structural descriptions of human consciousness across contemplative traditions, philosophy, and modern psychology, providing a representation-level perspective for analyzing and guiding model behavior.