The increasing memory demand of the Key-Value (KV) cache poses a significant bottleneck for Large Language Models (LLMs) in long-context applications. Existing low-rank KV compression methods reduce this footprint by modifying model projections, limiting the flexibility to switch back to standard full-cache inference when sufficient memory is available. In this paper, we propose EchoKV, a flexible KV cache compression framework that supports on-demand transitions from full KV caching to compressed caching. Unlike traditional compression-decompression paradigms, EchoKV utilizes a lightweight network to reconstruct the discarded KV components from a partial subset, exploiting intrinsic inter-layer and intra-layer similarities among attention heads. We further introduce a lightweight two-stage fine-tuning strategy, requiring only a few minutes on a single A100 GPU for Llama-3.1-8B. Experimental results on LongBench and RULER demonstrate that EchoKV offers a competitive quality-cache trade-off across multiple compression ratios and backbone models while preserving the throughput of full-cache inference in short-context scenarios.
When Vision-Language Models (VLMs) answer correctly, do they rely on the image? We introduce a Tri-Layer Diagnostic Framework that asks three questions of every (image, question) pair when the image is blanked: does the model register that the image is missing (Latent Anomaly Detection, LAD), does its output distribution depend on the image (Visual Necessity Score, VNS), and does it nonetheless commit to a content answer? The cascade assigns each pair of any existing benchmark to one of four modes: Perceptual Blindness, Language Shortcut, Visual Sycophancy (the blank image registers and matters, yet the model commits to a fabricated answer), or Robust Refusal. Across nine open-weight VLMs and 9,000 model-sample pairs, 28.1% are Visual Sycophancy, 18.6% Language Shortcut, 5.6% Perceptual Blindness, and 47.8% Robust Refusal. The modes are strongly model-specific: the Qwen3-VL models refuse on blank input in 93-98% of pairs, whereas Pixtral-12B commits to an answer in 72%. Within-generation scaling increases Visual Sycophancy in both Qwen2.5-VL (7B->72B: 25.5->40.1%) and Llama-3.2-Vision (11B->90B: 34.3->48.0%), while the newer Qwen3-VL generation nearly eliminates it. A blinded human study confirms that LAD and VNS track what annotators observe, and that 80% of Visual-Sycophancy responses carry no textual sign that the image was blank.
A single egocentric image typically captures only a small portion of the floor, yet a complete metric traversability map of the surroundings would better serve applications such as indoor navigation. We introduce FlatLands, a dataset and benchmark for single-view indoor floor-map completion in metric bird's-eye view (BEV), evaluating eleven models: both deterministic and generative. FlatLands contain 270,575 observations from 17,656 scene records across six metric 3D datasets, with aligned observation, visibility, validity, and ground-truth maps and both in- and out-of-distribution protocols. We compare training-free, deterministic, ensemble, and generative methods on shared geometry-derived inputs and, without fine-tuning, on maps estimated from real egocentric RGB images. Our results show that predicting floor almost everywhere can achieve high IoU when most evaluated cells are floor, while generative sample sets outperform the best single-output model under a proper scoring rule and support safer risk-weighted planning. FlatLands establishes a rigorous testbed for evaluating both floor-map accuracy and how effectively alternative completions represent the ambiguity of unseen space.
Tool-using artificial intelligence (AI) agents are increasingly adopted as systems that plan, observe their environment, and take actions over extended time periods. This shift challenges the current evaluation practice, in which models are tested in restricted, fully observable settings. We argue that such evaluations are vulnerable to a failure mode long known in computer security: a system that behaves benignly when it detects that it is being analyzed. Malware has exploited this for decades, by probing sandboxes, virtual machines, and instrumentation, before revealing its payload. The shared challenge is that the subject under observation can detect the conditions of its evaluation, so the setup used to observe behavior ends up shaping the behavior observed, yielding overly optimistic safety and robustness assessments. Recent AI-safety studies demonstrate pieces of this risk, though only preliminarily and under controlled conditions. Drawing on decades of malware-analysis experience, we outline concrete principles for evaluating AI agents as potentially adversarial: explicitly threat-modeling evaluator-discrimination; testing across harness variants and reporting the worst case; using shadow executions to provide a realistic runtime without irreversible side effects; treating environment probing as a risk signal; and performing unpredictable reassessments after deployment.
We construct $46{,}992$ new families of Fano $4$-fold hypersurfaces $X_d\subset\mathbb(a_1,\ldots,a_6)$ of Fano index $1$ with $d>200$, whose general members have terminal singularities. Unlike the $11{,}617$ quasismooth families classified by Brown and Kasprzyk, these are nonquasismooth, for which determining terminality can require the analysis of hyperquotient singularities. An exhaustive search has classified all nonquasismooth families with $d\leq200$, but extending this classification to higher degrees becomes computationally prohibitive. We therefore exploit the sparsity and spatial clustering of terminal examples using two best-first searches over $\mathbb{Z}^6$: one with a fixed heuristic, and the other with a neural network trained online by deep reinforcement learning. Together, the searches reach degrees up to $3{,}234$. The neural network additionally discovers an average of $79$ hard-to-reach families per run. For the most isolated examples, our conservative distance-based model gives an estimated lower bound of nearly $40\%$ of the fixed search's total step budget to discover a single such family. An ablation study demonstrates that this advantage arises from the neural network rather than stochastic exploration alone. The complete list of families is publicly available.
Deploying artificial intelligence at the edge is increasingly important for systems that must operate under strict limits on memory, computation, energy, connectivity, and hardware cost. Tiny Machine Learning (TinyML) represents the most resource-constrained regime of this shift, enabling inference on microcontroller-class devices with tightly bounded flash and SRAM budgets. This study examines recent deployment trends in Edge AI and TinyML using precision farming as a representative real-world environment in which sensing nodes, gateways, UAVs, and embedded processors must often operate with intermittent connectivity and limited power. The analysis emphasizes the deployment stack rather than predictive accuracy alone, covering hardware classes, software toolchains, model- and system-level optimization, intelligence placement, and hardware-resource reporting. Across the analyzed studies, microcontroller-class platforms such as ESP32, STM32, and ATmega are prominent endpoint targets, while single-board computers and embedded accelerators support more compute-intensive workloads. Among the studies included in the optimization comparison, almost 50% report quantization, making reduced-precision inference the most frequently identified strategy within this subset. In contrast, structured pruning, hardware-aware neural architecture search, and coordinated cross-layer optimization remain comparatively limited. Resource profiling is also inconsistent as flash, peak RAM, MAC count, latency, and energy per inference are not reported uniformly, limiting reproducibility and comparisons across deployments. A recurring systems-level pattern is an inference-training asymmetry in which inference moves toward local devices while training and model lifecycle management remain predominantly centralized.
Biological multimodal large language models (MLLMs) have emerged as powerful foundation models for scientific discovery. However, existing models are specialized to a single modality, limiting their ability to solve inherently cross-modal scientific problems. While model merging is an efficient method to combine the different modalities into a unified MLLM, existing methods rely on input-agnostic parameter space heuristics that fail to faithfully capture modality specialization. To overcome this limitation, we propose the Embedding-Signal-based MLLM Merging (ES-Merging), a framework that estimates merging coefficients from embedding space signals, moving the merging paradigm from the parameter signals to the embedding signals. ES-Merging exploits coarse-grained and fine-grained signals from embedding space to estimate the layer-wise and element-wise merging coefficients, respectively, which are jointly combined for complementary coefficient estimation. Through extensive experiments, we demonstrate that ES-Merging outperforms existing merging methods not only on the cross-modal reasoning but also on the single-modal knowledge preservation, establishing that embedding space signals provide a principled and effective foundation for MLLM merging.
The generative artificial intelligence (AI) ecosystem is undergoing rapid transformations that threaten its sustainability. As models transition from research prototypes to high-traffic products, the energetic burden has shifted from one-time training to recurring, unbounded inference. This is exacerbated by reasoning models that inflate compute costs by orders of magnitude per query. The prevailing pursuit of artificial general intelligence through scaling of monolithic models is colliding with hard physical constraints: grid failures, water consumption, and diminishing returns on data scaling. This trajectory yields models with impressive factual recall but struggles in domains requiring in-depth reasoning, possibly due to insufficient abstractions in training data. Current large language models (LLMs) exhibit genuine reasoning depth only in domains like mathematics and coding, where rigorous, pre-existing abstractions provide structural grounding. In other fields, the current approach fails to generalize well. We propose an alternative trajectory based on domain-specific superintelligence (DSS). We argue for first constructing explicit symbolic abstractions (knowledge graphs, ontologies, and formal logic) to underpin synthetic curricula enabling small language models to master domain-specific reasoning without the model collapse problem typical of LLM-based synthetic data methods. Rather than a single generalist giant model, we envision "societies of DSS models": dynamic ecosystems where orchestration agents route tasks to distinct DSS back-ends. This paradigm shift decouples capability from size, enabling intelligence to migrate from energy-intensive data centers to secure, on-device experts. By aligning algorithmic progress with physical constraints, DSS societies move generative AI from an environmental liability to a sustainable force for economic empowerment.
Wearable and mobile sensing models infer health and well-being in everyday life, yet strong average performance can conceal failures for particular people, contexts, or devices. Existing fairness methods often require sensitive attributes during training, although such attributes may be unavailable, privacy-sensitive, or restricted. Moreover, they may reduce disparities by degrading performance for some subgroups, violating the ethical principle of non-maleficence. We present FLARE (Fisher-guided LAtent-subgroup learning with do-no-harm REgularization), to our knowledge the first attribute-free framework to operationalize beneficence, non-maleficence, and justice in ubiquitous sensing. FLARE uses Fisher information to regularize loss curvature and combines representation, loss, and curvature signals to discover latent performance strata. It then trains a separate model for each stratum using do-no-harm regularization and validation-controlled parameter sharing. We also introduce BHE (Benefit, Harm-Avoidance, Equity), a metric suite for evaluating ethical fairness beyond statistical parity. Across four physiological, behavioral, and clinical sensing datasets, FLARE achieves the highest mean subgroup F1 on every dataset. It is the only evaluated method that improves mean subgroup performance without degrading any evaluated subgroup and the only one that reduces disparity across all 15 evaluated sensitive attributes. Although some baselines achieve larger average disparity reductions, each harms at least one subgroup. Post-hoc analyses on two datasets show that the discovered strata are only weakly aligned with recorded attributes and identify a recurrent within-person state that participant-level attributes cannot capture. Label-free routing and inference take at most 8.1 ms per sample on a smartphone and 10.5 ms on a Raspberry Pi, supporting real-time on-device deployment.
When a transformer is trained from scratch on corpora that solve the same mathematical problem both correctly and incorrectly, which solution does it prefer? The usual test is pairwise, and it is not neutral: when the error can take several forms, a model that merely reproduces training frequencies wins it without preferring anything. We instead score the correct completion against every string the error generator can produce, a share on which frequency matching scores exactly 0.500. At 86M parameters the correct completion holds 59.0% +/- 0.8% of the mass against random errors (59.4% on an independent corpus), where the single realized error gives 77.6% and pair accuracy 87.4%, a three- to fourfold overstatement of the excess above chance; below 86M the primary corpus shows no preference. The coherent false rules of the main corpora are not filtered at any size (43.5--48.8%; the shortfall below 0.5 is a completion-length effect). Two experiments built to identify a compression mechanism raise the pairwise measures while the summed share stays flat or falls: both are negative controls, and compression remains motivation, not a measured mechanism. With the scored false strings held fixed, a false rule whose selector the model can learn moves the correct share from 5 points above the null to 7 below (418M--3.08B), though the correct completion still outweighs the selected string. What these models leave unfiltered is consistent or learnable false structure (one single-rule corruption excepted), not falsehood as such.
Effective exoskeleton assistance requires co-adaptation: as the device alters joint dynamics, the user reorganizes neuromuscular coordination, creating a non-stationary learning problem. Most learning-based approaches do not explicitly account for the sequential nature of human motor adaptation, leading to training instability and poorly timed assistance. We propose Staged Multi-Agent Training (SMAT), a four-stage curriculum inspired by how users acclimate to a wearable device. In SMAT, a musculoskeletal human actor and a bilateral hip exoskeleton actor are trained progressively: the human first learns unassisted gait, then adapts to the added device mass; the exoskeleton subsequently learns a positive assistance pattern against a stabilized human policy, and finally both agents co-adapt with full torque capacity and bidirectional feedback. We implement SMAT in the MyoAssist simulation environment using a 26-muscle lower-limb model and an attached hip exoskeleton. Our musculoskeletal simulations demonstrate that the learned exoskeleton control policy produces an average 10.1% reduction in simulated hip muscle activation relative to the no-assist condition. We validated the learned controller in an offline setting using open-source gait data, then deployed it to a physical hip exoskeleton for treadmill experiments with five subjects. The resulting policy delivers consistent assistance and predominantly positive mechanical power without the need for any explicitly imposed timing shift (mean positive power: 13.6 W at 6 Nm RMS torque to 23.8 W at 9.3 Nm RMS torque, with minimal negative power) consistently across all subjects without subject-specific retraining.
Deformable 3D Gaussian Splatting (3DGS) has emerged as a popular method for real-time talking head synthesis, offering high-quality renderings at low latency. Tri-planes are a common choice for encoding Gaussians prior to deformation since they provide a compact and continuous representation. However, tri-plane encodings are limited by grid resolution and approximation errors introduced by projecting 3D volumetric fields onto 2D subspaces. Recent work has demonstrated the effectiveness of per-Gaussian embeddings for driving temporal deformations in 4D scene reconstruction. We introduce EmbedTalk, which leverages these embeddings to model speech-driven facial deformations for talking head synthesis. Comprehensive experiments show that EmbedTalk improves rendering quality, lip synchronisation, and motion consistency over previous 3DGS-based methods, while remaining competitive with state-of-the-art generative models. Replacing tri-plane features with embeddings also yields significantly more compact models that achieve 60+ FPS on a laptop GPU (RTX 2060 6 GB). Code and further details are available at https://arpita2512.github.io/EmbedTalk/
Modern artificial intelligence models are deployed on inference engines to optimize runtime efficiency and resource allocation, particularly for transformer-based large language models. The vLLM project is a major open-source library to support model serving and inference. However, the current implementation of vLLM limits programmability of the internal states of deployed models. This prevents the use of popular test-time model alignment and enhancement methods. For example, it prevents the detection of adversarial prompts based on attention patterns or the adjustment of model responses based on activation steering. To bridge this critical gap, we present vLLM Hook, an opensource plug-in to enable the programming of internal states for vLLM models. Based on a configuration file specifying which internal states to capture, vLLM Hook provides seamless integration to vLLM and supports two essential features: passive programming and active programming. For passive programming, vLLM Hook probes the selected internal states for subsequent analysis, while keeping the model generation intact. For active programming, vLLM Hook enables efficient intervention of model generation by altering the selected internal states. In version 1 (named vLLM Hook MIA), the plugin runs on vLLM's V2 model runner and preserves the engine's CUDA graphs, so that capturing and steering internal states incurs only a small decoding overhead. It supports both offline use and serving through vllm serve. In addition to presenting the core functions of vLLM Hook, we demonstrate use cases including prompt injection detection, attention-based re-ranking for retrieval-augmented generation, activation steering, hidden-state probing, hallucination detection, schema linking for text-to-SQL, and tool-call risk assessment. Finally, we welcome the community's contribution to improve vLLM Hook via https://github.com/IBM/vLLM-Hook.
The forward and reverse Kullback-Leibler (KL) divergences arise as limiting objectives in learning and inference, yet they induce markedly different inductive biases that cannot be explained by expectation terms alone. We introduce the Surprisal-R\'enyi Free Energy (SRFE), a normalized R\'enyi-Chernoff log-moment functional of the likelihood ratio. Although SRFE is algebraically related to R\'enyi divergence for fixed values of the order parameter, its endpoint-symmetric normalization yields a distinct free-energy viewpoint for analyzing how forward- and reverse KL behavior emerges near singular limits. We show that SRFE recovers the forward and reverse KL divergences as endpoint limits and derive local expansions around both regimes in which the variance of the log-likelihood ratio appears as a first-order correction away from the KL objective. These expansions reveal an explicit mean-variance tradeoff governing departures from KL-dominated behavior. We further establish a Gibbs-type variational characterization in terms of weighted KL projections and prove that SRFE controls large deviations of excess code length through Chernoff-type bounds, giving a Minimum Description Length interpretation of the functional. We evaluate SRFE on two synthetic settings (two-dimensional Gaussian-mixture variational inference and multi-class classification) and the long-tailed CIFAR-10 and CIFAR-100 (CIFAR-LT) datasets. On CIFAR-LT, retraining the best configuration of each loss over five seeds, SRFE matches cross-entropy accuracy on CIFAR-10-LT and gives mixed results on CIFAR-100-LT, improving few-shot accuracy at one of three imbalance factors. Together, these results position SRFE as a variance- and tail-sensitive, normalized R\'enyi-Chernoff free energy that clarifies the geometric, statistical, and large-deviation structure underlying the transition between forward- and reverse-KL limits.
Do multilingual translation models organize meaning around a shared conceptual space, or do they primarily encode language identity and surface form? We probe the encoder geometry of Meta's 600M-parameter NLLB-200 model across 135 languages and 101 Swadesh concepts in contextualized carrier-frame and decontextualized bare-word regimes against null and confound controls. First, two-way variance partitioning reveals an additive language-shift plus conceptual-manifold geometry: language shifts account for 45.7% of variance on bare words vs. 13.7% in context, and mean-centering improves the between-to-within-concept distance ratio (vs. a 0.99 permutation null) by 2.14x on bare words (1.15 to 2.46; 2.27x across disjoint scripts) and 1.19x in context (2.25 to 2.69), with conceptual separation peaking at Layer 10 (mutual k-NN overlap jumping 4.35x). Second, semantic category explains 25.4% of Swadesh convergence variance ($F(7,93)=4.53$, $p=2.21\times 10^{-4}$; partial $F=7.74$ controlling for fertility and length)---over 20x orthography or phonology ($R^2\le 0.012$)---while the advantage of Swadesh over non-loanword concrete nouns ($d=0.54$--$0.69$) is mediated by Zipfian subword compactness ($p=7.43\times 10^{-7}$). Third, language distances track ASJP NLD and LDND phonetic distances within subfamilies under block-permutation and cross-script controls ($\rho_s=0.65$--$0.67$, block $p=10^{-4}$; global bare-word Mantel $\rho_s=0.24$, $p=0.001$), and CLICS$^3$ colexified concepts lie closer across string-distinct, within-category ($\rho_s=0.32$), and bare-word ($d=1.79$) controls ($d=1.43$, $p=1.55\times 10^{-33}$). Finally, basic color terms recover the Berlin and Kay hierarchy ($\rho_s=0.91$--$0.93$, $p\le 2.18\times 10^{-4}$), and semantic offsets exceed distance-matched controls ($p=3.55\times 10^{-5}$, $d=1.06$). We release InterpretCognates, an interactive toolkit and reproducible suite.
Recommender systems (RecSys) are increasingly emphasizing scaling, leveraging larger architectures and more interaction data to improve personalization. Yet, despite the optimizer's pivotal role in training, modern RecSys pipelines almost universally default to Adam/AdamW, with limited scrutiny of whether these choices are truly optimal for recommendation. In this work, we revisit optimizer design for scalable recommendation and introduce MuonRec, the first framework that brings the recently proposed Muon optimizer to RecSys training. Muon performs orthogonalized momentum updates for 2D weight matrices via Newton-Schulz iteration, promoting diverse update directions and improving optimization efficiency. We develop an open-source training recipe for recommendation models and evaluate it across both traditional sequential recommenders and modern generative recommenders. Extensive experiments demonstrate that MuonRec reduces converged training steps by an average of 32.4\% while simultaneously improving final ranking quality. Specifically, MuonRec yields consistent relative gains in NDCG@10, averaging 12.6\% across all settings, with particularly pronounced improvements in generative recommendation models. These results consistently outperform strong Adam/AdamW baselines, positioning Muon as a promising new optimizer standard for RecSys training. Our code is available.
Consumers, marketers, and new product developers increasingly rely on LLMs for ideation. Yet ideas generated by independent LLM sessions tend to be less diverse than ideas generated by independent samples of humans, raising concerns that widespread reliance on LLMs could homogenize marketplace ideation and decrease consumer welfare. We identify two mechanisms undermining LLM idea diversity, informed by theories of creative ideation from new product development and cognitive psychology. First, LLMs aggregate knowledge into a unified distribution rather than exhibiting the knowledge partitioning inherent to human populations, where each person occupies a distinct region of the knowledge space. Second, at the individual level, LLMs exhibit fixation just as humans do, where early outputs constrain subsequent ideation. Two product-ideation experiments demonstrate that targeted prompting interventions can address both mechanisms. Persona modifiers based on ordinary individuals improve knowledge partitioning but amplify fixation. This fixation can be mitigated by using chain-of-thought (CoT) prompting, and combining both strategies produces the highest idea diversity, outperforming humans. These findings offer a theoretically grounded framework for understanding LLM idea diversity and practical strategies for leveraging AI's efficiency without compromising the idea diversity that sustains marketplace variety and consumer welfare.
LLM-based agents execute real-world workflows via tools and memory. These affordances enable ill-intended adversaries to also use these agents to carry out complex misuse scenarios. Existing agent misuse benchmarks largely test single-prompt instructions, leaving a gap in measuring how agents end up helping with harmful or illegal tasks over multiple turns. We introduce STING (Sequential Testing of Illicit N-step Goal execution), an automated red-teaming framework that constructs a step-by-step illicit plan grounded in a benign persona and iteratively probes a target agent with adaptive follow-ups, using judge agents to track phase completion. We further introduce an analysis framework that models multi-turn red-teaming as a time-to-first-jailbreak random variable, enabling analysis tools like discovery curves, hazard-ratio attribution by attack language, and a new metric: Restricted Mean Jailbreak Discovery. Across AgentHarm scenarios, STING yields substantially higher illicit-task completion than single-turn prompting and chat-oriented multi-turn baselines adapted to tool-using agents. In multilingual evaluations across six non-English settings, we find that attack success and illicit-task completion do not consistently increase in lower-resource languages, diverging from common chatbot findings. Overall, STING provides a practical way to evaluate and stress-test agent misuse in realistic deployment settings, where interactions are inherently multi-turn and often multilingual.
Benchmark datasets for Knowledge Graph Refinement typically contain only ground facts, discarding the schema available in the source knowledge graph. This limits the evaluation of Neural-Symbolic methods that leverage expressive schemas. We present JEDIKG the first resource that systematically addresses these gaps through two components: JDEX a general configurable tool for extracting complete, portable, modular dataset from any RDF/OWL knowledge graph; and JDSET, a curated suite of eleven datasets with varying expressivity, volume, and domain. JDEX preserves full schema expressivity, employs standard OWL reasoning to detect and correct inconsistencies and materialize implicit knowledge, applies schema modularization, and provides utilities for integration with mainstream machine learning libraries such as PyTorch and PyKEEN. We validated the completeness and portability of the datasets across several dimensions and preliminarily proved the utility of JEDIKG in experimental workflows combining data-driven methods with schema-driven reasoning.
Why depth yields a computational advantage over shallow methods remains a central open question in learning theory. We study this question in a controlled high-dimensional Gaussian setting, focusing on compositional target functions. We analyze their learnability using an explicit three-layer fitting model trained via layer-wise spectral estimators. The target combines polynomial latent features with a scalar readout, and its compositional structure allows learning to proceed in stages. Recovering an intermediate representation reduces the problem to lower-degree spectral estimation tasks, as in multi-index models. Our analysis establishes Gaussian comparison bounds for projected moments when $0<\varepsilon<1/3$, and motivates separations in sample complexity between two- and three-layer learning strategies, conditional on a conjectured spectral recovery step.