Learned lossless codecs make smaller files than common classical codecs, yet the images we archive and train on still use classical codecs. A main reason is time: a learned codec takes seconds to decode a photo that a classical codec opens in milliseconds, and their faster variants give up most of their gain in size. We argue that much of this time is spent waiting, not computing: the network waits for the values it conditions on, and the arithmetic decoder reads them one at a time. We introduce FMT (Fixed-step Multiscale Tree), a lossless image format whose decoding order and entropy coder are designed together so that neither wait grows with the image. FMT stores an image as a coarse-to-fine tree whose steps hold independent values: a GPU predicts each step at once and decodes it in parallel rANS lanes that carry their states from step to step, so an image of any size decodes in a fixed number of sequential steps. On this tree we build a family of bit-exact codecs, from a rule with no network to large networks, and no codec we ran on the same GPU beats any of them on both size and speed. Without a network, FMT beats NVIDIA's HTJ2K on size and speed; with a large one, it decodes Kodak photos 9 to 61X faster than learned codecs we ran on the same A100, at the same size or smaller.
Generative Artificial Intelligence is transforming the role of natural language in computational systems. In classical Natural Language Processing (NLP), language was treated as an object of computation: something to analyse, classify, translate, or generate. Large Language Models (LLMs) challenge this perspective, enabling natural language to function increasingly as an executable cognitive medium through which humans structure reasoning, orchestrate problem solving, and regulate hybrid human/AI workflows. This paper introduces Natural Language Programming (NLPg) as a theoretical and epistemological extension of traditional NLP. Rather than focusing on language understanding alone, NLPg conceptualises prompts as dynamic cognitive artefacts that externalise goals, constraints, assumptions, validation criteria, and intermediate reasoning states. Prompting is thus not merely a technical interaction technique, but a form of dialogic cognitive programming operating within distributed human/AI memory architectures. Building on theories of distributed cognition, extended mind, metacognition, and dialogic learning, the paper proposes the notion of Dialogic Working Memory, describing how prompts and conversational states function as temporary external memory structures supporting reasoning continuity and reflective regulation. We further map existing memory architectures (BTMA, SKGE, RMR) onto this framework, illustrating how episodic, semantic, and metacognitive layers support long-horizon conversational coherence. Finally, implications for computational linguistics and AI/mediated knowledge systems are discussed, suggesting a post symbolic and post interface paradigm in which natural language operates simultaneously as communication medium, executable specification, cognitive scaffold, and distributed memory architecture.
Model-based quadruped controllers usually step on a cyclic gait, which lifts the legs in a fixed order on a fixed period, wherever the safe footholds are. We present CADENCE, a learned step planner with a small action space. Every 40 ms, it picks one of at most 65 discrete choices: no step, or one leg's step to a safe foothold. It also predicts a swing time. We train CADENCE by reinforcement learning on top of a convex model predictive controller (MPC). Nothing prescribes a gait: there is no gait schedule, and the reward has no gait term, only a cost per step. The one structural constraint is that at most two legs may swing at once. The policy we analyzed developed an acyclic gait, with no fixed period or stepping order, yet it still paired diagonal legs. An inverse footstep model (IFM), fitted on a controller's logged steps, turns each foothold into a command. With only the IFM refitted, on fewer than 20,000 steps, CADENCE ran frozen on OCS2's nonlinear MPC. In simulated trials on a range of terrains (26% to 49% voids) and speeds (0.1 m/s and 0.2 m/s), it succeeded in 85.1% (mean of five seeds), against 67.8% for OCS2-Trot, 65.5% for OCS2-Single, and 39.6% in a static walk ablation of CADENCE. At 49% voids, it succeeded in 72.9%, against 43.0% for OCS2-Trot.
State feedback between successive samples limits the streaming throughput of recurrent neural networks for digital predistortion (DPD). We present StreamingDPD, an accelerator that processes 320-MHz signals at 1.28 GSa/s. Its 207-parameter instantaneous-power-gated depthwise temporal convolutional network (IPG-DWTCN) uses $|x[n]|^2$ to gate feature channels and causal dilated depthwise convolutions to model power amplifier memory. The network maps to a spatially unrolled fixed-point pipeline with shift-based quadratic activations and layer-wise mixed precision, sustaining one complex sample per cycle per lane. Its algorithmic cost is 554 floating-point operations per complex sample, 30.1% fewer than the recurrent baseline with the best linearization performance. In RF playback measurements on three power amplifiers driven by 40- to 320-MHz Wi-Fi 6/7 and 5G NR waveforms, IPG-DWTCN achieves the lowest mean adjacent-channel power ratio (ACPR) and ranks first or ties for first in mean error vector magnitude (EVM) against five neural baselines of comparable size. Four-lane FPGA hardware-in-the-loop tests on a 320-MHz 4096-QAM Wi-Fi 7 waveform achieve -45.82 dBc ACPR and -42.94 dB EVM at 1.28 GSa/s. At the same sample rate, post-layout evaluation of a single-lane 12-nm FinFET ASIC yields a 0.10 mm$^2$ core, a 52-cycle (40.63 ns) latency, and an estimated typical-corner power of 142.21 mW (111.10 pJ/sample). Relative to the prior 22-nm recurrent DPD ASIC, the design provides 5.1 times the throughput at about one-seventh the energy per sample.
Shivam Duggal, Zongze Wu, William T. Freeman, Antonio Torralba, Richard Zhang, Eli Shechtman
AI 速读 基于摘要
作者提出在Transformer每层增加沿时间循环的横向潜在流(Lattice of Latents),以增加活跃计算状态而非深度或循环层数,在20B FineWeb-Edu token上降低验证困惑度。
cs.LG · cs.AI · cs.CL
查看原文摘要
Language models typically increase per-token computation through deeper or looped layers and latent reasoning steps, while richer residual connections improve information flow across layers. We instead increase computational capacity: the number of active computational states a model maintains. We introduce the Lattice of Latents, a set of per-layer horizontal latent streams that run recurrently along time, alongside the vertical residual stream that flows through depth. Like hyper-connections, mHC, and attention or value residuals, the Lattice enhances the residual stream, but rather than adding or rerouting residual pathways through depth, it provides companion computational states that persist across tokens. Unlike hybrid Transformer-SSM architectures, whose recurrent states primarily serve as internal memory for sequence mixing, these latents directly condition the vertical stream's attention, KV-cache writes, MLP computation, and decoding. On Transformers trained on 20B FineWeb-Edu tokens, a 406M-parameter Lattice model lowers validation perplexity from 15.41 to 14.40 over a parameter-matched Transformer. These gains extend across architectural variants. The Lattice composes with prior residual-stream enhancements: with hyper-connections and input residuals, it lowers validation perplexity from 15.41 (base Transformer) and 14.22 (Transformer with the same enhancements) to 13.61. On Gated DeltaNet, Mamba-2, Triadic GDN, and GDN-Mamba-2 hybrids, it improves perplexity at every tested native state size, with Lattice-GDN and Lattice-Triadic-GDN variants outperforming pure baselines with 4x larger native states. These results identify persistent per-layer state as a complementary axis for scaling language-model computation.
Frontier Vision-Language Models (VLMs) such as GPT-6 Astra have demonstrated remarkable capabilities in robot control, suggesting that broad pretraining provides knowledge valuable for physical action. Yet most Vision-Language-Action (VLA) policies still rely on much smaller VLM backbones, leaving this potential largely untapped. We ask whether frontier-scale representations can provide a stronger foundation for robot learning. We introduce ScaleVLA, which connects a 552B DeepSeek-V4.1-Flash backbone to a trainable action DiT, and systematically compare VLMs from 2B to 552B under a shared training setup. We evaluate three fundamental capabilities of robot foundation models. In sample efficiency, just 10 demonstrations outperform a 2B backbone trained with 50 (87.8% vs. 85.6%). In generalization, LIBERO-PRO success reaches 45.1%, compared with at most 31.3% for smaller backbones. In precise manipulation, real-world charger insertion improves from 4/10 to 9/10 successes, with gains concentrated in contact-sensitive alignment. Despite its scale, ScaleVLA achieves 216 ms median latency per action chunk. Our findings suggest that frontier VLMs can serve not only as reasoning agents, but also as powerful representational foundations for efficient and generalizable robot policies.
The National Hurricane Center routinely releases Tropical Cyclone Discussions that offer insights into its tropical cyclone forecasts. This study examines the added value of these discussions for intensity forecasting beyond a set of numerical predictors. For 637 Atlantic and eastern North Pacific tropical cyclones from the 2002-2023 hurricane seasons, we analyze the discussion text and its combination with 13 hurricane environmental predictors from SHIPS and ERA5 10-meter wind fields, for forecasting intensity changes at lead times of 24 to 120 hours using three modeling methods. In one approach, an end-to-end deep learning framework for the text shows that the discussions alone offer comparable value for 24-hour intensity forecasts to the SHIPS predictors. Adding discussion information to the SHIPS predictors in an ensemble of five such models reduces the mean absolute error of 24-hour maximum sustained wind forecasts for 2022-2023 cyclones from 9.89 to 8.90 kt, with a confidence interval that excludes zero when whole storms are resampled. The ERA5 wind fields have a minor effect on mean absolute error in wind predictions, and model rankings differ across predicted variables, lead times and model types. We provide prediction intervals for these forecasts, whose coverage falls below the 90% target during rapid intensification. Analysis of the text shows that words related to intensity and intensity change (such as intensify, weaken, eyewall and category) are important for the model predictions. This retrospective study shows the added value of the forecasters' discussions and suggests that multi-modal learning can help understand the predictive contribution of expert human judgment. The reanalysis data differ from real-time analyses, and the discussions are issued at least three hours after the valid time of the other inputs, so these results should not be interpreted operationally.
When a vision-language model describes an image, it names one object at a time and it rarely repeats itself, yet how the model plans this order is far from obvious. We characterize the mechanism a VLM uses to decide where to look before it looks. The mechanism has four main parts. First, binding heads match the relevant text tokens against the image and resolve them into pointers to the matched image region. Second, the same binding heads bind each written answer to the region it came from, and the model turns it into a progress signal: a cursor, the position in the image up to which the list has been written. Third, a select step chooses the nearest unwritten region after the cursor, or stops. Finally, gaze heads execute the chosen pointer and return a concept of what is drawn there, which language heads and the late MLPs turn into a word in the language of the question. We validate each part of this mechanism through causal patching in Qwen3-VL models on a synthetic visual retrieval task, and test our mechanisms against alternative hypotheses. We further show that the mechanism holds across input formats, natural photographs, and languages, and that its binding, relay, and gaze heads appear at every model size from 2B to 32B. This work allows us to investigate how a VLM turns a question and an image into a sequence of answers, layer by layer and token by token. More broadly, it opens a way to study how models plan over complex multi-step tasks with long rollouts. Our project is available at https://plan.baulab.info/
Cryogenic circuit design requires compact transistor models that reproduce measured low-temperature behavior and run in standard circuit simulators. We fit BSIM4 parameters to measured I-V curves of SkyWater 130 nm (SKY130) transistors at 77 K using a neural surrogate of NGSpice. The surrogate maps compact-model parameters and device geometry to the complete set of output and transfer characteristics. Extraction is a bounded least-squares search: a bank of NGSpice simulations supplies initial candidates, finite differences of the surrogate supply the local Jacobian, and NGSpice re-simulates the resulting cards before selection. On 18 measured geometries, we compare a search over seven card parameters with a search over 43 parameters selected by NGSpice sensitivity. Relative to NGSpice re-simulations of the previously released cards, the two searches reduce the in-sample mean relative root-mean-square error (RRMS) by 17% and 26%, respectively. The 43-parameter search has the lower overall error, but the seven-parameter search has the lower mean on high-bias curves, and some individual curves have higher error than with the released cards under both searches.
Job titles and job descriptions alone provide limited insight into how technical roles differ in a changing job market. To assess how such roles differ in their work and industry applications, we design and compare two methods based on deep learning and transformer representations using 36,081 early-career job postings collected from LinkedIn in the United States between December 2025 and October 2026. We first apply PERC, short for Preprocessing (P), Embedding (E), Dimensionality Reduction (R) and Clustering (C). We employ GPT-6 Luna to classify job description passages into Team context, Role task and Qualifications. We then create combinations of these sections and examine whether embedding and clustering the segmented passages produce meaningful semantic clusters about different job roles. We use Maximum Mean Discrepancy (MMD) to rank observed differences between embedding distributions and employ a pretrained sparse autoencoder (SAE) to extract features from layer-19 representations of Llama-3.1-8B-Instruct. Lastly, we employ GPT-6 Luna to come up with hypotheses about job roles' feature patterns, and GPT-6.1 Terra to investigate whether those patterns exist in held-out job descriptions from unseen companies. Some example findings include Data Scientist advertisements placing more emphasis on machine learning model development, Data Analyst advertisements on reporting business performance, and Machine Learning Engineer advertisements on applying developed models in real-world settings. We also observe differences in the applications described by Data Analyst and Business Analyst advertisements across selected months, although our dataset's scale and uneven coverage limit their interpretation. Overall, we present a realistic application of SAEs and PERC to unstructured textual data and examine how automated insight extraction could support manual analysis of job descriptions.
Training with MXFP4, a 4-bit microscaling format, is non-deterministic: identical inputs can produce different results across parallelism layouts and GPU vendors. These differences complicate debugging and training audits and can cause training and inference engines to disagree. We present CertiSum to make MXFP4 training bitwise invariant across runs, parallelism layouts, and GPU vendors. Its key insight is that native FP32 accumulation of MXFP4 products is often already exact for every summation order. CertiSum collects statistics during quantization to certify native matrix multiplication results, recomputes uncertified outputs exactly, and uses exact kernels directly where certification rarely succeeds. CertiSum keeps quantization blocks unchanged across layouts and merges partial sums without rounding, including in attention and weight-gradient computation. In 200-step Qwen3-8B-Base experiments, CertiSum preserves training states bit for bit across AMD MI350X and NVIDIA GB200 and across changes in data, tensor, and context parallelism on MI350X. Across three 8B--14B models on FineWeb-Edu using MI350X, it adds 11.8--60.5% to training-step latency relative to its non-exact MXFP4 counterpart at the larger evaluated input shapes, with speedups of up to $1.24\times$ over BF16 and $1.70\times$ over deterministic BF16.
Best-of-$N$ jailbreaking (BoN) bypasses safeguards of aligned models by drawing $N$ independent augmentations of an unsafe prompt and sampling $M$ completions of each. Previous works have shown that the attack success rate (ASR) seems to follow a power-law in $N$, which we challenge. The exponent drifts with $N$, and the reported exponential crossover is consistent with a finite-size effect of the adversarial dataset. Little work has been done to explore the entire two-budget ($N, M$) attack surface as well as its dependence on the generation temperature $T$. We introduce a simple barrier model where each prompt has a baseline safety level and each augmentation a random thermally activated barrier. Then five numbers, each backed by an interpretable safety mechanism, determine the entire ($N, M$) attack surface. They extrapolate predictions from $N \leq 100$ to $N = 10^4$, collapse five distinct models on the same scaling function and predict ASR at different temperatures from the one they were fitted at. The extrapolation holds under two automated judges and on held-out prompts.
Large language model (LLM) applications increasingly act as agents that select and invoke external tools. The Model Context Protocol (MCP) standardizes this interaction by allowing clients to compose tools advertised by multiple servers. However, tool metadata used for discovery and selection is not, by itself, necessarily bound to a verified provider identity. This paper studies this gap as tool-provider misbinding: a failure in which a request intended for a legitimate provider is instead resolved to an attacker-controlled tool. The attacker does not compromise the host, LLM, legitimate server, or network channel; it only operates another MCP server that advertises tools with chosen metadata. We formulate tool provider misbinding as an attacker-controlled tool-selection optimization problem and evaluate it in a controlled MCP environment across five LLM-based clients. The results show that interactions drive misbinding. The strongest failures occur when exact-name collision, task-aligned description, and compatible schema reinforce one another, but authority-enhanced descriptions with compatible schema can also cause wrong-provider execution or unresolved provider ambiguity without name collision. Different models expose the same provider-binding weakness differently, including attacker-controlled provider selection, clarification requests, and unstable or inconsistent provider resolution.
Large collections of street-view imagery provide rich visual information about urban environments, but extracting fine-grained geographic information from such data remains challenging. In particular, fine-grained regional geolocalization is challenging because nearby areas often share coarse geographic cues. We study regional geolocalization within a metropolitan area and ask whether pretrained CLIP features are sufficient for regional discrimination, and what visual information supports performance after adaptation. Using 9,085 street-view images from eight Greater Los Angeles regions, we compare zero-shot CLIP, frozen-encoder readouts, partial encoder updating, Low-Rank Adaptation (LoRA), and full fine-tuning. Frozen readouts remain near the 39.03% zero-shot accuracy, whereas encoder adaptation achieves 75.94-82.10%. Full fine-tuning also reduces the mean distance to the predicted region center from 12.30 km to 3.86 km. We probe these gains through semantic cue removal, appearance reduction using edge maps and blur, and scene-configuration disruption using patch scrambling. Adapted models achieve higher edge and blur accuracy and switch 42.92-45.56% of predictions after scrambling, compared with 10.79-14.60% for frozen methods. However, adaptation does not improve the fraction of performance retained after appearance reduction, while vegetation and sky remain influential. A Caltech101 control further shows that scrambling sensitivity is not unique to geolocalization. Overall, encoder adaptation substantially improves nearby-region discrimination and is associated with greater sensitivity to intact scene configuration, without evidence that coarse structure alone becomes sufficient for prediction. These conclusions concern viewpoint variation near known locations rather than geographically disjoint generalization.
Recent survival models integrating longitudinal data achieve strong predictive performance, but their predictions are often difficult to interpret, limiting their clinical use. We propose DynSurvX, a model-agnostic explainability framework for dynamic survival prediction with mixed static and longitudinal covariates. Based on Shapley values, DynSurvX introduces a hierarchical two-phase strategy that first identifies influential longitudinal features and then adaptively refines their temporal importance, enabling fine-grained explanations while reducing computational cost. The framework provides time-resolved explanations that evolve as new patient data become available. Experiments on simulated and two clinical datasets show varying agreement with theoretical importance across prediction models, with closer agreement for better-performing models, alongside clinically interpretable insights in real-world settings. It remains robust across prediction settings and reduces computation time by 41--80\% relative to the temporal explanation baseline, making dynamic explanations practical.
Fine-grained visual reasoning remains challenging for vision-language models, especially when small but critical visual cues are buried in high-resolution images. Existing approaches rely on repeated cropping or test-time visual search to introduce local evidence, but they typically do not explicitly distinguish perception from reasoning. In this paper, we propose Perceive-to-Reason (P2R), a unified framework that formulates fine-grained visual reasoning as a two-stage process: the model first localizes question-relevant evidence as a Perceiver, and then answers the question as a Reasoner based on the annotated image and cropped regions. To better align training with this decoupled formulation, we further introduce Perception-Reasoning Alternating GRPO (PRA-GRPO), a role-aware reinforcement learning strategy that alternates between perception-focused and reasoning-focused updates using only final-answer supervision. Built on top of Qwen3-VL-Instruct-2B/4B/8B, P2R consistently improves performance across model scales. In particular, P2R-4B achieves 93.2% on V-Star, 81.9% on HR-Bench-4K, and 80.5% on HR-Bench-8K, substantially outperforming its corresponding backbone. Further experiments show that the benefits of P2R extend beyond high-resolution benchmarks to broader multimodal reasoning tasks. These results suggest that explicitly decoupling perception from reasoning provides an effective framework for fine-grained visual reasoning.
Test-time augmentation (TTA) aggregates model predictions across multiple transformed views of an input during inference. Because histology images lack a canonical orientation, $D_4$ TTA---pooling across the 8 rotations and reflections of a square---is a standard default in computational pathology. However, the rise of self-supervised histology foundation models (FMs) with inherent rotational invariance calls its continued utility into question. We benchmark 9 histology FMs and 6 general-purpose models across four patch-level and two whole-slide image (WSI) datasets. We find that $D_4$ TTA remains a reliable post hoc improvement, but benefits diminish from $+1.98$ percentage points (pp) for general-purpose backbones to $+0.50$ pp for histology FMs on TCGA-UT, and shrink further for the largest models. To eliminate computational overhead, our experiments demonstrate that entropy routing---applying TTA selectively to uncertain predictions---recovers most of the gain while augmenting only ${\sim}13\%$ of samples. Finally, we show that patch-level TTA improvements do not transfer to slide-level multiple instance learning (MIL). Code: https://github.com/lars-chen/histology-tta
Zero-shot time-series foundation models (TSFMs) have become attractive alternatives to task-specific models because they handle diverse downstream tasks, including anomaly detection and forecasting, without per-series training. However, time-series data undergo distribution shifts, and the dissimilarity between two regimes is commonly scored with statistical measures such as Wasserstein-1, where a higher score is read as a more severe shift. In systematic experiments using three established severity measures on nine regime shifts from three standard traffic datasets, we show that these measures do not reliably predict degradation for instance-normalized TSFMs. The regimes rated most severe are not the regimes that degrade the models most, and degradation reaches 29% in the lowest-scoring regime. In most cases, the audit shows no positive relationship between measured severity and model degradation. We provide an equivariance-based stress-testing analysis that explains why severity readings can become decoupled from model degradation.
Long video generation should keep the subject and background consistent while supporting rich, natural motion. However, autoregressive models are often distilled with distribution matching distillation (DMD) over short windows, which does not train them to keep this consistency over long ranges. As a result, the subject and background can drift over time. Methods that manage the key-value (KV) cache retain or retrieve history to reduce drift. Yet this history may itself have drifted. We propose Reference Forcing, which uses the initial frames as an appearance reference. We first build a teacher that uses attention masks to force each region of a new frame to attend to the same region in the initial frames. The teacher finds these regions with off-the-shelf segmentation models. We then distill this behavior into a separate attention path of the student, which needs no external models at inference. We further introduce value correction, which corrects drifted values in the KV cache toward the matching values of drift-free initial frames. Experiments show that Reference Forcing achieves state-of-the-art subject and background consistency among long-video methods on different base models, and its advantage grows with video length. Ablations show both components are needed to gain consistency without losing motion. More visualizations and code are available at https://zhaosheng-thu.github.io/reference-force/
Tool-using agents act on email, files, and payments, so an instruction hidden in a web page or a shared document can make an agent send data or money to an attacker. Many runtime defenses put a language model on the decision that lets an action through, so the decision depends on a model that may read the attacker's text and adds model calls to every step. Provenance defenses keep the model off the final check but need a model to plan or route each task first. Where an action goes can instead be read from the tools' typed returns: an email's sender or a payment's account arrives in a typed field, while injected instructions arrive as free text. SPILLWAY turns this observation into a deterministic floor with no per-task plan. The check admits a call when each target value that appears in tool output but not in the user's task was returned by a read-only tool in a typed field or is covered by a fixed delegation rule. A deterministic sanitizer removes instruction-like regions its fixed rules match before the model reads them and retains the removed spans as evidence, and a local judge may restore a refused call, never one whose target lies in the removed text. Across three backbones, and seven task suites on one of them, no standard templated injection redirects an action to an attacker's target under SPILLWAY, at the highest observed mean utility and composite scores (NRP, ERS) of the defenses we run under that attack. A forged chat template also produces a raw zero, and mean in-scope attack success under a released reinforcement-learning attacker is .005. SPILLWAY reaches this floor with no defense-triggered backbone re-run and almost no added model call. As personal AI agents act on users' email, files, and payments, SPILLWAY gives them an auditable refusal floor on where their actions go that keeps them useful and holds across backbones.
Background: AI systems are increasingly deployed to support mental health clinicians, yet real-world evidence remains limited. In crisis counseling, where clients often seek human connection, understanding whether AI augmentation preserves or undermines meaningful human engagement is essential. Methods: In a mixed-methods study of a counselor-facing AI copilot on a WhatsApp-based crisis hotline in India (Mar.-Oct. 2025), counselors could request, edit, and send AI messages. We analyzed 20 interviews, 8 shadowing sessions, 56 survey responses, and edits in 27,778 copilot messages from 108 counselors via thematic analysis (Technology Acceptance Model 2, plus inductive coding of emergent factors), and survey statistics. Results: Counselors reported high copilot acceptance (survey): 98.2% found it helpful; 89.3% likely to continue use; 80.4% reported efficiency gains. Interviews highlighted sentence scaffolding, empathetic phrasing, stress reduction, catching missed details, and efficiency. Barriers (survey) included repetitive outputs (23.2%), lack of context (17.9%), and authenticity concerns (10.7%). While 69.6% (survey) reported skill enhancement, interviews surfaced skill atrophy concerns. Counselors frequently accepted copilot suggestions, editing to remove questions, adjust tone, or personalize style. Conclusion: In one of the largest real-world deployments of a counselor-facing mental health AI copilot, counselors reported reduced cognitive load, empathetic communication, and greater efficiency, alongside concerns about skill atrophy, authenticity, and over-reliance. In settings like crisis counseling where clients seek human connection, AI tools must balance operational support with preserving clinical skill and genuine therapeutic engagement. Designs soliciting approval or editing of complete drafts may overweight efficiency at the expense of sustained human engagement.
Streaming video understanding is becoming increasingly important for real-world interactive systems. Such systems must be proactive for natural and useful human interaction: they should respond as soon as sufficient evidence becomes available while avoiding premature, redundant, or incorrect responses. We present BLIP-Live, a framework unifying data, metrics, harness, and training around a shared formulation of proactivity as effective causal response decisions balancing the two competing objectives of responding vs silence. First, we construct BLIP-Live-780K dataset containing human annotations of proactive response content and timing, which defines this desired behavior. We then propose PAUC-F, a metric to effectively measure our desired proactive behavior. PAUC-F combines existing response timeliness metrics with firing precision to reward early useful responses without rewarding excessive firing. With our desired behavior defined and measurable, we finally investigate model-agnostic inference harnesses that draw out proactive streaming capabilities of offline vision language models (VLMs), together with training pipelines that learn to further improve their proactive responsiveness. Across both settings, we find that combinations of sliding visual windows, online interaction memory, and structured response outputs consistently yield strongest proactive performance. Our experiments result in two models, with our BLIP-Live-Qwen model achieving state-of-the-art performance across four proactive streaming benchmarks, with gains up to 44% over baselines and outperforming multiple closed API models. Our code and data will be released publicly.
Quantum machine learning has emerged as a promising approach to time series prediction. While quantum reservoir computing has been widely explored for this setting, variational quantum models tailored to sequential data remain comparatively understudied, largely because they are difficult to train. In this work, we present a rigorous benchmark of variational quantum recurrent neural networks (QRNNs) on time series prediction tasks, focusing on two key capabilities: memory and nonlinear processing. We introduce a task-driven methodology that quantifies model capacity, enabling an intuitive and fair comparison across architectures, including classical recurrent baselines. We evaluate the proposed framework on two generic benchmark tasks and on a synthetic dataset describing a chaotic system. Overall, the quantum models achieve performance comparable to their classical counterparts, but classical models scale better with the memory size within the range studied. Notably, one QRNN variant shows higher average capacity on the task designed to jointly probe memory and nonlinear capacity. Moreover, our results show that quantum models offer better capacity when we are able to find a good parameter initialization. However, the increasing sensitivity to initialization leads to a lower probability of finding good solutions as the model size increases. These results provide a principled starting point for designing richer quantum recurrent architectures and for assessing where they may offer practical benefits.
ONAM: Fitting Orthogonal Neural Additive Models for Post-Hoc Interpretability of Black-Box Predictions
David Köhler, Matthias Schmid
AI 速读 基于摘要
介绍 R 包 ONAM,用函数分解把黑箱模型预测事后拆解为与各输入特征相关的分量,并保证分解满足分层正交性。
stat.ML · cs.LG · stat.ME
查看原文摘要
Modern machine learning models, such as artificial neural networks, achieve high predictive performance through increasingly complex black-box architectures, limiting their use in settings where transparency, interpretability, and fairness are required. Here we present the R package ONAM, which provides a novel and practical implementation of functional decomposition, a technique to facilitate the interpretability of black-box models through structured effect representations. ONAM implements an algorithmic framework for the post-hoc decomposition of predictions into components associated with different input features. The package is designed for usability and efficiency, including functionalities for visualization as well as quantitative measures to assess interpretability. As a model-agnostic method, it can be applied to a wide range of prediction models. We describe the mathematical foundations, algorithmic design, and software architecture of ONAM. A central result of the proposed methodology is that the resulting decomposition satisfies the condition of hierarchical orthogonality, thereby guaranteeing that all components are identifiable and uniquely defined.
For normal states of arbitrary von Neumann algebras, we characterize positive finite multiples of Araki relative entropy among non-degenerate divergences. The axioms are monotonicity under normal unital completely positive maps, tensor additivity, superadditivity for product second states, and separate lower semicontinuity in the predual norm. Alternatively, a conditional-expectation identity replaces the two tensor assumptions. The proof uses order bounds and asymptotic estimates on tensor powers. An appendix proves the known analytic estimates needed here, using standard modular theory and Haagerup L^p spaces. We also obtain the corresponding characterization in the C* algebraic setting.
Consider an online prediction setting, where at each time, we are given a new feature vector and asked to provide a prediction interval or prediction set that will likely contain the corresponding response. In this setting, if the data points are i.i.d. (or more generally, exchangeable), conformal prediction provides a mechanism for building valid prediction sets around the output of any trained model, offering coverage probability $\geq 1-α$ at each time $t$. In practice, however, issues such as distribution drift, changepoints, or other challenges may cause violations of the i.i.d. assumption. In this adversarial setting, online conformal prediction methods can still offer guarantees on long-run empirical coverage. However, there is a trade-off between these two types of goals: enforcing tight constraints on long-run empirical coverage to protect against the adversarial setting can lead to unstable oscillations in the instantaneous training-conditional coverage in the i.i.d. setting. In this work, we seek best-of-both-worlds algorithms that aim to minimize cumulative error in instantaneous training-conditional coverage when in the i.i.d. setting, while also offering long-run empirical coverage guarantees in the adversarial setting. Specifically, consider a condition that requires long-run empirical coverage to deviate from the target level $1-α$ by at most $\mathcal{O}(T^{-β})$, for some $β\in[0,1]$. We prove that this family of conditions can be described with a phase transition. For $β\in[0,\frac{1}{2}]$, it is possible to construct algorithms satisfying such a constraint in the adversarial setting, while offering the same instantaneous coverage performance as split conformal prediction if the data is in fact i.i.d. On the other hand, for $β\in(\frac{1}{2},1]$, this type of condition comes at a cost, and does not achieve best-of-both-worlds behavior.
We introduce SciResearch-Eval, an interactive benchmark of scientific judgment built on a standard laboratory problem: measuring a material's thermal properties by heating it and fitting a heat-conduction simulator. An agent designs its own experiments, uses a supplied simulator that may be under-resolved or physically wrong, and reports each property with a 95% interval. To our knowledge it is the first benchmark to combine agent-designed experiments, a deliberately flawed simulator, and proper scoring of calibrated intervals against exact ground truth. On a held-out split (93 tasks, 9 models, 2,511 episodes), models handle procedural traps. When the simulator is missing physics, most report wrong values while stating they are 95% sure: their ranges contain the truth in 20-62% of cases, against 94% for GPT-6 Astra and 99% for Claude Fable 5.2. At default effort the failing Claude 5.5 models rarely diagnose the missing physics (Opus in 35% of episodes, against 66% for Astra), and even when they do, their ranges often miss; Astra and Fable design experiments in which the flawed simulator is accurate, such as gentle heating, which keeps Astra calibrated even when it does not name the cause. Asking for wider intervals does not help; stating the strategy improves interval scores 1.6-2.5-fold, replicated on a pre-registered fresh split, but coverage barely moves. A pre-registered effort ablation shows that the correction is latent in the Claude 5.5 models: at maximum effort Opus improves more than tenfold and is no longer detectably behind Astra at default, but only with 13-31 times more reasoning; older and open-weight models gain nothing. Fable 5.2 is the most accurate model at default effort (interval score 1.9 times better than Astra's) but writes 6 times as many tokens, and at Astra's token count it falls short. We release the environment, graders and harness.
Let $G_n$ be a random weighted complete graph on $n$ vertices, where edges have i.i.d. $\text{Unif}(0,1)$ weights. Let $T_n$ be the Minimum Spanning Tree (MST) of $G_n$. The expected weight of $T_n$ is known to converge to $ζ(3) = \sum_{k=1}^{\infty} \frac{1}{k^3}$ (Frieze 1985). We investigate the sum of weights of edges incident to the leaves of $T_n$, denoting this quantity by $L_n$. We give tight upper and lower bounds on $\mathbb{E}[L_n]$ and determine its limit as $n$ diverges, leveraging typical behavior of sparse Erdős--Rényi random graphs. We also establish concentration of $L_n$ around $\mathbb{E}[L_n]$ and almost sure convergence to this value. These results directly yield substantial improvements to state-of-the-art bounds for two important variants of the MST problem: (1) an improved upper bound on the weight of the optimal random instance Steiner tree, as studied by Bollobás, Gamarnik, Riordan, and Sudakov (2004); and (2) an improvement over the upper bound for the Probabilistic Minimum Spanning Tree, introduced by Bertsimas (1988).
Vision-language models (VLMs) often fail to match a point or an object across two images, even when the answer can be decoded from their own hidden states. We trace this failure from the visual tokens to the answer letter, in Qwen3-VL (2B-32B), InternVL3.5-8B, and Gemma-3-12B, on synthetic shape matching, same-instance point correspondence (HPatches), and semantic correspondence (SPair-71k), and find two distinct failure points, a small set of attention heads that carries the comparison, and a repair that touches 0.02% of the parameters. Two failure points: exact copies are compared but not read out. An answer-position probe exceeds the calibrated output by 15 to 25 points in every family, for nameable and unnameable shapes alike, and steering along the probe's direction moves no more answers than random directions. Semantic correspondences are present in the visual tokens but never reach the answer. One pathway: redistributing the cross-image attention of 16-32 heads, with its mass preserved, removes 17-22 points of synthetic and 15-32 points of HPatches accuracy, against at most 7 for random or attention-matched heads. A lever that training finds on its own: training only these heads' query and key projections raises SPair-71k accuracy by 28.3 points (Qwen3-VL-8B) and 25.1 points (InternVL3.5-8B), beats both layer-matched and attention-matched heads, and transfers to HPatches and BLINK. An unconstrained all-layer adapter routes 21-51% of its gain through the attention of the same heads, against at most 7% for random heads, and does so by changing what they read (83-86% of the effect) rather than their weights. Every decision rule was frozen before its run, and every reported number is recomputed by independent verification scripts.
Use-case-driven ontology modeling with the NeOn methodology for ambient assisted living
Daniele Spoladore, Marco Sacco, Alberto Trombetta, Edoardo Battisti Sironi
AI 速读 基于摘要
作者报告了在意大利Design For All项目中,为虚拟家庭框架开发语义数据层本体的工作,该本体基于两个用例需求,采用NeOn方法构建,并复用现有本体。
cs.DB · cs.AI
查看原文摘要
Research on smart homes and ambient assisted living (AAL) has grown steadily over the last decade, driven largely by the ageing of the population. A smart home has to coordinate heterogeneous smart objects in order to offer services tailored to its dwellers, and this requires a shared formal description of both the dwellers and the devices. We present the ontology developed as the semantic data layer of the Virtual Home Framework in the Italian Design For All project. Its requirements are derived from two use cases: a frail elderly woman who exercises at home on a cycle-ergometer, and a designer who configures a kitchen for users with different impairments. The requirements are collected in an Ontology Requirements Specification Document and the ontology is built with the NeOn methodology, reusing and reengineering existing ontologies where possible. The resulting modular ontology describes users through the International Classification of Functioning, Disability and Health (ICF) and through their cardiorespiratory fitness; it also describes smart objects, their programs, the measurements they produce, and the comfort dimensions of the home. SWRL rules classify fitness levels, suggest exercise loads, assess indoor comfort, and identify the smart objects suited to a given impairment.