Papers

On-policy self-distillation (OPSD) adapts a language model by distilling guidance from a frozen teacher on trajectories sampled from the student. Its effectiveness, however, depends critically on the quality of those trajectories. We show that when student rollouts drift from target trajectories, conditioning the teacher on off-target prefixes substantially weakens its task-relevant supervision. Controlled prefix-corruption experiments expose this failure mode, which we term rollout-conditioned signal degradation. To address this problem, we propose a unified training framework that separates two complementary supervision pathways. The first retains rollout-conditioned distribution matching, providing guidance on states the student actually visits. The second applies supervised cross-entropy on canonical ground-truth contexts, avoiding the incompatibility of imposing target tokens on erroneous rollout prefixes. Token-level rollout-target alignment is used to adapt the strength of the canonical-context anchor, emphasizing it during cold start and relaxing it as rollout quality improves. Experiments across multiple model scales, two task families, and general-reasoning benchmarks show that the proposed approach improves task acquisition over OPSD while preserving general capabilities, resulting in a more favorable empirical plasticity-stability trade-off. These findings identify context quality as a central bottleneck in on-policy self-distillation and demonstrate the value of separating rollout-conditioned guidance from canonical supervision.

Drifting models are a recent class of one-step generative models that evolve the model distribution during training using a predefined sample-based drift field. Although they avoid iterative inference, their kernel-based drift fields induce frequency-dependent training dynamics: In the linearized regime, each Fourier mode of the density residual decays at a rate determined by the kernel spectrum, leading to slow recovery of fine-scale structure. We propose Second-Order Drifting Models, which lift drifting dynamics into phase space by augmenting generated samples with artificial velocity variables. We show that the resulting density perturbations obey accelerated second-order dynamics in Fourier space, connecting drifting models to the celebrated Nesterov acceleration from optimization theory. This provides a principled mechanism for mitigating the spectral stiffness of first-order drifting while preserving one-step inference. We derive a practical semi-implicit training algorithm and evaluate it on synthetic distribution matching, sequential data generation, and robotic control. Across these settings, the second-order drifting model improves convergence behavior and achieves competitive or superior performance over first-order drifting baselines.

Lipschitz bandits admit near-optimal regret $\widetilde O_d(T^{(d+1)/(d+2)})$ with little memory under full adaptivity, or with few batches under unrestricted memory. We characterize the minimax expected pseudo-regret over $T$ rounds with $W$ bits of memory and at most $B$ batches in dimension $d$. For every memory budget $W$, we prove a lower bound of order $T^{\frac{d+2}{d+3}} (1+(B-1)W)^{-\frac{1}{d(d+3)}}$. When $W\gtrsim_d\log(eT)$, algorithms with fixed batch boundaries attain, up to logarithmic factors, the larger of this bound and the optimal $B$-batch regret with unrestricted memory. The analysis separates fine-scale comparisons from the regional information needed to allocate their samples at low regret. Attaining $\widetilde O_d(T^{(d+1)/(d+2)})$ regret requires both $B=\Omega_d(\log\log T)$ and $(B-1)W=\widetilde{\Omega}_d(T^{d/(d+2)})$. With one bit of memory, minimax regret is $\widetilde\Theta_d(TB^{-1/(d+2)})$ under adaptive batch boundaries while $\Theta_d(T)$ under fixed boundaries.

We apply Marchenko-Pastur (MP) random matrix theory to pre-trained attention weights in order to separate each projection matrix into a random-like bulk and a set of spectral outliers. We validate this decomposition causally: zeroing the MP-identified outliers (signal) in Mistral-7B drives HellaSwag, MMLU, and PIQA close to random-chance performance, whereas zeroing a count-matched subset of bulk singular values causes smaller but non-negligible degradation. Across 11 pre-trained transformers we identify five recurring patterns: spectral outliers encode a dominant component of the learned structure; Q projections carry the most outliers; V projections under grouped-query attention lack a clean signal/noise separation; entry-level outliers form structured row-bands in Q and column-bands in O; and specific residual-stream dimensions persist as band outliers across layers in K and O. We close by outlining how these observations could inform parameter-efficient fine-tuning and structured pruning.

Large language models (LLMs) increasingly read long inputs in the agentic era, from whole documents and codebases to conversations across many turns. Their inference memory is then dominated by the key-value (KV) cache, the stored attention keys and values of every token the model has read and generated. Because the cache grows with context length and is re-read in full at every generated token, a longer context means more GPU memory. To reduce this cost, most existing methods compress the KV cache by lowering every stored value to the same low precision, a technique known as quantization. They can push this to nearly two bits per value, but rarely further, because quality drops sharply at this 2-bit cliff: four levels are too few for the cache's outlier-heavy values, where a few large entries consume the levels and collapse the rest into noise. A natural remedy is to spend more bits on the channels (feature dimensions) that matter and fewer on the rest, but the raw cache offers no handle: its channels are strongly correlated, so none stands out as more important. Our analysis shows that this handle appears once the cache is rotated into a coordinate system computed from its own statistics, removing these correlations. There, a small fraction of channels carries almost all the information, and spending the budget on those few is far more accurate than spreading it evenly. Guided by this analysis, we develop SPECTRA, a training-free, drop-in codec that re-encodes the cache into this coordinate system and concentrates the bit budget on the channels that carry the signal. On Llama-3.1-8B and Qwen2.5-7B over long-context benchmarks, SPECTRA is near-lossless at 4x compression, competitive at 8x where uniform quantization has collapsed, and reaches up to 12x, pushing usable compression past the 2-bit cliff so the same GPU holds longer contexts and larger batches.

Mixture-of-Experts (MoE) models have outgrown accelerator memory, and offloading expert weights to host memory is now standard. This makes expert cache management an attractive lever: a policy that raised the hit rate would cut expert traffic per token. Evaluating that is a measurement problem, and we find the measurement fragile. With a trace-driven, event-atomic simulator over three MoE models (40, 64, 128 experts), we isolate three evaluation axes that change conclusions, not just numbers. Replay semantics: under a fused-event traffic contract, an inconsistent per-access replay inflates recency-based policies by 27-29% while leaving frequency-based and static ones within 4%, inverting the policy ranking. Workload contamination: probe sets using one instruction template per category produce verbatim-identical generation prefixes; a matched-pair rendering intervention moves the measured early-window effect by 19.4-31.9 points and reverses which workloads look most cache-friendly. Operating regimes: normalized miss fractions do not transfer across models, so the per-step expert union relative to per-layer capacity must be reported – yet permuting only the temporal order of an identical event stream moves the offline-optimal gap from 44.9% to 30.8%, so it is not sufficient. Corrected, a stable gap to the offline optimum remains (44.2-45.9% over 13 frozen workload compositions). A forced-admission oracle attributes 84.3-96.6% of it to knowing which resident expert is used furthest in the future. A causal next-use predictor, used as an eviction rule, recovers -11.4% of the gap; it picks an optimal victim 3.4% of the time, against 2.4% for a random resident block and 20.6-22.1% for LRU and LFRU. Our position is narrow: in our evaluated settings a large offline-optimal gap substantially overstates the gains recovered by representative lightweight causal mechanisms.

Robot demonstration datasets used to train vision-language-action policies can contain a subtle but harmful failure mode: trajectories that are behaviorally correct but paired with the wrong language instruction. We study post-hoc auditing of these Instruction-Trajectory Mismatches (ITMs). Unlike failed rollouts, ITMs often look plausible, and can corrupt the language-behavior mapping learned by the policy. We propose Multimodal Probabilistic Fusion (MMPF), a training-free auditing framework that treats each modality as an expert, estimates a task-label distribution from local neighborhood agreement and global prototype similarity, and then fuses modalities with predictive-entropy weighting in a product of experts. Across LIBERO benchmarks with injected instruction mismatches and noisy real-robot data, MMPF achieves the strongest overall ITM detection and label correction accuracy. We also show that auditing improves most downstream policy learning in settings where language is needed to disambiguate the task. We demonstrate in real robot experiments that our method can achieve improved policy performance and show the trade-off of filtering demonstrations compared to relabeling.

Recent advances have highlighted the potential of machine learning, particularly Large Language Models (LLMs), for analyzing and optimizing programs. We present the first application of program embeddings from LLMCompiler—an LLM massively pretrained on intermediate representation (IR) code—to representative program analysis and optimization tasks. We generate program embeddings directly from source and IR code using a simple approach: split programs into chunks, independently embed each chunk with pretrained LLMs, and then aggregate the chunk embeddings into a single program embedding. Our experiments show that combining source and IR code embeddings achieves an error rate of 1.54% in algorithm classification, a 12% improvement over the current state-of-the-art, and a competitive accuracy on heterogeneous device mapping. These findings suggest that training a performance-aware LLM for embedding IR code might yield state-of-the-art results in code optimization tasks.

Mixture-of-Experts (MoE) models decouple parameter count from per-token compute, but deployment still requires hosting every expert in memory. Recent theory shows that experts whose router weights change least during fine-tuning can be pruned with provable accuracy preservation, yet the guarantee assumes full fine-tuning. We show that the signal can be elicited through a brief parameter-efficient adaptation. We fine-tune with a lightweight adapter, rank experts by the induced router change, and prune the least-changed experts in one shot. On Mixtral-8$\times$7B-Instruct, router-only LoRA trains 0.002% of parameters and retains 27.54% MMLU-Pro accuracy with half the experts removed, against roughly 16% for magnitude and random pruning. Signal quality improves monotonically with adapter size, reaching 28.76%, and declines as adaptation spreads beyond the router. Under their shared budget, IA3 reaches 28.04% while Houlsby reaches 25.39%. The criterion transfers to Qwen1.5-MoE fine-tuned for mathematical reasoning, retaining 49.7% mean accuracy over eleven benchmarks with half the experts removed. Structural pruning reduces memory by 49% and per-token latency by 37%. Lightweight router sensitivity therefore makes provably motivated, task-conditioned expert pruning practical at scale.

Clustering mixed tabular data requires a unified metric space to bridge the inherent heterogeneity between continuous numerical measurements and discrete categorical symbols. Traditionally, algorithms rely entirely on dataset-internal statistics to estimate categorical relationships, which confines the learned metric to empirical co-occurrences and ignores conceptually obvious yet statistically unobserved affinities. Although LLMs offer external world knowledge, applying their text-centric reasoning to highly abstract tabular concepts presents significant challenges. Bridging this modality gap to construct a semantically complete metric typically requires embedding LLMs into iterative metric learning loops to dynamically optimize cross-modality representations. This incurs intractable computational overhead, forcing a compromise between semantic enrichment and scalability. Therefore, we propose GRACE, an LLM-grounded framework for scalable mixed-data clustering. GRACE shifts semantic acquisition to the attribute-value level via a multi-perspective LLM querying strategy, mapping heterogeneous values into knowledge-informed descriptions. Crucially, this one-shot grounding extracts general-purpose semantic representations that embed heterogeneous attributes into a unified space, decoupling expensive LLM invocation from iterative optimization. Furthermore, GRACE cross-validates these external semantics against dataset-internal statistical evidence to ensure alignment with the dataset-specific cluster structure. Ultimately, GRACE matches the scalability of conventional statistics-driven baselines while achieving superior clustering accuracy and conceptual interpretability over 11 competing methods. The source code is available at https://github.com/develop-yang/GRACE-GRACE-A

Accurate, up-to-date income data at the sub-municipal scale is essential for social policy in middle-income countries, yet in Brazil it depends on a costly decennial census whose intercensal gap recently exceeded a decade. We test whether the composition of crowd-sourced Google Maps Points of Interest (POIs) can serve as a high-frequency, low-cost proxy for household income across the 26,625 census sectors of the municipality of Sao Paulo. Using a theoretically motivated set of POI categories retrieved from Google Places, we represent each sector by its POI counts, decompose these high-dimensional, sparse features with principal component analysis (PCA) and non-negative matrix factorization (NMF), and train a sweep of regression models to predict census-derived income. Under a data leakage-aware spatial validation design the best model (NMF with gradient boosting) attains a held-out R^2 of 0.65, with performance stable across feature-extraction methods. Interpretable decompositions reveal which POI types carry the income signal. These results suggest that commercial, crowd-sourced geospatial data can complement conventional income statistics during intercensal periods, and we discuss extensions toward multidimensional poverty and the capabilities framework.

Improving sample efficiency remains a core challenge in reinforcement learning (RL), especially in real-world settings like robotics, where data collection is costly. This challenge is pronounced in visual RL, where high-dimensional inputs often obscure learning signals. While prior work in visual RL has focused on algorithmic solutions, such as better dynamics models or exploration strategies, recent advances in state-based RL show that architectural design alone can lead to significant gains in sample efficiency. This raises an important question: Can these architectural principles transfer to visual RL? In response, we introduce V-Simba, a simple yet effective visual RL architecture inspired by the Simba architecture from state-based RL. Built on top of Soft Actor-Critic (SAC) with data augmentation, V-Simba modifies the architecture by adding normalization layers to stabilize training and using pointwise convolutions to reduce computation. Despite its simplicity, V-Simba matches or outperforms the state-of-the-art methods across the DMC, Adroit, and Meta-World benchmarks, while being more computationally efficient than DrQ-v2. We make our code publicly available at https://github.com/DAVIAN-Robotics/V-Simba.

Multimodal emotion recognition often treats self-reported labels as reliable supervision while overlooking self-report unreliability and cross-modal conflict. We propose CONFER, a graph-based conflict-aware evidence negotiation framework for weakly supervised multimodal emotion recognition. CONFER represents each modality expert as a node with a predictive belief, boundary-based uncertainty, and runtime reliability estimated from historical out-of-fold performance and current-sample uncertainty. Uncertainty-aware compatibility and reliability-directed asymmetric edge weights govern iterative message-passing negotiation, followed by peer-supported prediction readout. Conflict reduction, residual disagreement, and mean modality uncertainty further characterize three regimes—Consensus, Dissent, and Ambiguity—for sample-specific weak-label calibration. We evaluate CONFER on AMIGOS, MAHNOB-HCI, and DEAP under subject-dependent 10-fold and strict leave-one-subject-out (LOSO) protocols. CONFER achieves competitive performance, reaching 0.873 accuracy on AMIGOS-V and 0.854 accuracy on MAHNOB-V under strict LOSO evaluation. Further analyses show larger negotiation gains on high-conflict samples and improved robustness to weak-label corruption, indicating that cross-modal conflict provides useful information for both directional modality coordination and supervision-reliability estimation.

Preferential Bayesian optimization (PBO) optimizes objectives accessible only through pairwise user comparisons. The standard approach fits a Gaussian process surrogate for observed pairwise comparisons (PairwiseGP) using the Laplace approximation and selects queries with the Expected Utility of Best Option (EUBO) acquisition function. EUBO queries new candidates at each step, producing pairs that share no candidates with previous queries. Each such pair forms an isolated component in the comparison graph, removing one degree of freedom from the likelihood Hessian and making it rank-deficient. This deficiency is structural and cannot be resolved by changing the surrogate modeling approach. Existing approaches to remedy this issue either waste query budget by forcing comparisons to stay connected, or apply uniform regularization that also perturbs directions already well-constrained by the observed comparisons. We propose KappaSharp that enables a diagonal correction to the Hessian to reduce its condition number, with larger corrections where the prior uncertainty is higher. The correction is only applied in the model fitting step, not query selection. An adaptive variant of KappaSharp is also presented that activates the correction only when the surrogate is confident about recent comparisons, avoiding unnecessary corrections when the problem is well-conditioned. On 11 benchmarks (5–20 dimensions), including a 16-dimensional controller tuning problem in plasma medicine, Adaptive KappaSharp outperforms the standard PBO baseline, with up to +10.9% ($p{=}0.003$).

Multi-turn Reasoning-and-Acting (ReAct) agents accumulate growing trajectories of reasoning, tool calls, and observations. Their key-value (KV) caches grow accordingly, increasing memory use and attention cost during model inference. Existing KV cache compression methods reduce these costs by evicting states with low attention scores. However, low attention in the current turn does not imply future irrelevance, as temporarily inactive information may become important later. Snapshot-based eviction methods therefore do not explicitly distinguish temporarily dormant information from information that appears to have completed its role. In this paper, we present CommitKV, which identifies KV lifecycles through commit transitions. Specifically, CommitKV first divides completed agent events into token pages and compares each eligible page's deletion effect before a tool-call commit and after the commit's returned observation has been incorporated. Based on these paired measurements, CommitKV distinguishes dormant pages from high-to-low completion candidates. It then applies a greedy joint test, accepting candidates for retirement only when their combined post-commit effect remains bounded. Finally, at a later compression checkpoint, accepted pages are excluded, a bounded set of pages awaiting post-commit measurement is protected, and the remaining KV states are retained within the cache budget using the same token indices for keys, values, and absolute positions. These mechanisms ensure that CommitKV can distinguish dormant information from information that has completed its observed role and can be safely removed. Experiments on various benchmarks show that CommitKV reduces agent memory use, accelerates end-to-end inference, and achieves higher accuracy than existing KV cache compression methods.

Residual connections rely on a static residual pathway, and are essential for training deep neural networks. Hyper-connections (HC) increase the expressivity of residual routing by incorporating multiple residual streams and learning dynamic information flow, while manifold-constrained (mHC) variants stabilize training through doubly stochastic residual mixing. However, a generator-level bottleneck remains in existing methods: they use dense, unstructured generators for pre-branch aggregation, residual mixing, and post-branch redistribution, which results in parameter count growing rapidly with the number of streams. To address this issue, we propose \underline{T}ensorized \underline{E}fficient \underline{M}anifold-constrained \underline{P}arameterization for \underline{E}xpressive Residual \underline{R}outing (TEMPER), which represents these generators as multi-way tensors over the input-stream, feature, and output-stream modes, and parameterizes them using tensor networks. Such a structured low-rank formulation is shown to preserve token-dependent manifold-constrained routing interface while substantially reducing parameter growth. It also promotes interpretability and intuition, as: i) tensor ranks control the dimensionality of the learned routing subspace, with full ranks recovering dense routing; while ii) the generator approximation errors bound differences in routing logits and, consequently, in the routed-block outputs. Comprehensive experiments show that TEMPER matches or outperforms existing methods across language modeling and commonsense reasoning tasks, while requiring substantially fewer additional parameters. At eight residual streams, TEMPER achieves the best CORE score while using about $84\%$ fewer additional parameters than mHC, thus showing a stronger performance-parameter efficiency trade-off.

Confidence estimation for large language models (LLMs) aims to estimate the probability that a generated answer is correct, while calibration aligns these estimates with empirical accuracy. Prior work has shown that token probabilities are often overconfident, we investigate whether these readily available signals can nevertheless provide well-calibrated confidence estimation for mathematical question answering. We compare single-pass estimators, which reuse token probabilities from the original generation, with multi-pass estimators, which obtain additional confidence signals through verification or stochastic forward passes. While individual token probabilities can be highly saturated, we find that aggregating token probabilities over the full sequence captures small but consistent differences between correct and incorrect generations, yielding more informative confidence estimates. Multi-pass methods can yield calibrated confidence estimates. We study two such approaches: self-verification through re-prompting, including a lower-cost in-situ variant, and Monte Carlo Dropout, which derives confidence from variation across stochastic forward passes. We further evaluate two post-hoc calibration methods, Platt scaling and isotonic regression, both of which substantially reduce in-domain calibration error. However, their data efficiency varies with dataset difficulty, and the calibration mappings often transfer asymmetrically across datasets and models.

Running vision-language navigation fully onboard an aerial robot is hard, since grounding, planning, and control must share limited compute and a single-stage error is difficult to isolate in flight. End-to-end aerial policies fuse these stages into one network, giving up the observability and safety checks a modular stack keeps available. We propose VLN on the Fly, an onboard stack that keeps grounding, planning, and control as separate, inspectable stages. A quantized VLM grounds an instruction to a coarse image cell, depth lifts it to a 3D goal, a fast B-spline planner returns a feasible trajectory, and a pretrained reinforcement learning policy tracks it to motor commands across quadrotors. Across 15 onboard flights over three everyday referents in a controlled indoor volume, the stack reaches the target in 13 of 15 trials with 5.72 cm mean goal error and 39.3% average GPU utilization. In 6 additional cluttered-environment trials, the stack tracks collision-free trajectories under onboard perception gating.

We introduce StableEval Arena, a cost-aware benchmark framework for evaluating agentic AI systems on stablecoin peg-risk prediction. StableEval Arena evaluates LLM-backed agentic systems on diagnosing peg stress and forecasting deviations from the one-dollar peg over a hidden seven-day horizon, using leakage-safe historical replay with exchange price-volume data and market-context features. We report two complementary experiment blocks: a 120-case stress-enriched validation block and a 507-case natural-distribution full-arena evaluation block. Across six LLM-backed agent configurations and baselines, StableEval Arena measures prediction quality, calibrated-label behavior, structured-output reliability, latency, token consumption, and estimated inference cost. Rather than ranking agents by accuracy alone, the framework treats trustworthiness as a joint property of forecast quality, operational reliability, and computational cost. The results show a gap between protocol-following reliability and financial-risk reliability: agents reliably produce valid structured outputs at modest measured cost, but still miss most rare severe-stress and sustained-depeg cases. To support auditing and replication, we release the benchmark dataset on Hugging Face and the source code on GitHub.

Public health forecasts must respond to abrupt changes in surveillance data without over-extrapolating noise, reporting artifacts, or temporary trends. We evaluated autoregressive integrated moving average (ARIMA), random forest, and extreme gradient boosting (XGBoost) models using 190 weekly observations of publicly available Ontario COVID-19 case counts from January 2020 to October 2023. Rolling-origin time-series cross-validation preserved temporal order during model tuning and evaluation. Performance was assessed across three operating dimensions: responsiveness following selected turning points, forecast horizons of one to six weeks, and the amount of historical training data. We also developed Machine Learning and ARIMA Model Averaging (MLAMA), a non-negative performance-weighted ensemble with weights that vary by forecast horizon and responsiveness setting. Retrospective comparisons showed that ARIMA adapted rapidly after turning points but its normalized error increased at longer horizons. Random forest and XGBoost were less responsive initially but maintained more stable normalized error over longer horizons. For two-week forecasts at the end of the study period, training on the most recent data outperformed using longer historical periods, particularly for XGBoost. MLAMA achieved the lowest normalized mean absolute percentage error across most forecast horizons and ranked among the best-performing methods across responsiveness settings. These findings support selecting forecasting models according to operating conditions rather than relying on a single universally preferred approach. MLAMA provides a practical framework for combining complementary statistical and machine-learning forecasts. The accompanying Python package is currently maintained in a private repository while software validation and reproducibility testing are completed.

Sequence models are conventionally distinguished by their backbone, the mechanism that routes information across positions, such as attention or recurrence. This paper varies a choice that is prior to the backbone and shared by nearly all current models: the substrate, the number system in which the hidden state is represented together with the form of the map from state to prediction. The prevailing substrate is a real-valued state with an affine–softmax readout; we study a complex-valued alternative drawn from the mathematics of quantum theory, in which information is carried by the phases of the state and scores are quadratic Born forms. Prior work proved an idealized version of this substrate representationally stronger than any real model with a linear readout; we ask whether it also trains faster. Relaxing the two properties that block deployment, exact unitarity and the Born vocabulary readout, we instantiate it in the Mamba state-space model and an attention-based Transformer. At 253M parameters, matched to within $0.02\%$ and trained under one fixed protocol on three byte-level corpora, the complex models reach every measured validation loss in approximately one third (state-space) and one half (attention) of the optimization steps of their real counterparts. The two backbones then diverge. Once the learning-rate warmup ends, the state-space advantage continues to widen, from $0.321$ to $0.354$ bits per character on OpenWebText and from $0.368$ to $0.396$ on FineWeb, which an artifact of the warmup ramp would not do; the attention advantage instead decays toward zero on every corpus, and is therefore an effect of early training.

LLM recommenders for top-K item suggestion regularly emit titles outside the target catalog. Prior audits report a binary out-of-domain rate; none ask whether the model knew. We jointly audit hallucination rate (OOD@10) and verbalized-confidence calibration (ECE, Brier, reliability) for four zero-shot LLM recommenders from four independent vendors (Mistral Large, Llama-3.3-70B, GPT-OSS-120B, Claude Sonnet 4.6), not grounded or fine-tuned systems, across three catalogs (MovieLens-25M, Amazon Reviews 2023 Toys, Yelp Open Dataset), stratified by item popularity. Measuring catalog membership is itself the hard part: on identical outputs the reported rate moves by an order of magnitude with the string matcher used, and F1 cannot separate the candidates. We validate the instrument against 201 human judgments and select on net bias, where the adopted one is off by -0.040 against +0.144 for the common fuzzy rule. Hallucination is then strongly catalog-dependent (0.6-2.7% on MovieLens, 11.6-38.7% on Yelp, 49.3-61.0% on Amazon Toys). Each model holds a near-constant confidence level barely responsive to the catalog, while the catalog-hit rate swings 60 points, so the sign of the error is set by where a model's constant lands against a catalog's accuracy: 7 of the twelve cells are under-confident and 5 over-confident, all four under-confident on MovieLens, all four over-confident on Amazon Toys. We read this as an elicitation mismatch: "Just Ask" elicits a generic quality rating, not a catalog-membership probability. A conformal abstention threshold over verbalized confidence changes hallucination by at most 1.65 pp across four alpha levels, because the channel cannot separate correct items from hallucinations. We recommend that audits report calibration alongside OOD, validate the matcher producing the OOD number, and use catalog-anchored elicitation.

The current state-of-the-art (SOTA) deep randomized neural networks, such as deep Random Vector Functional Link (dRVFL) and ensemble deep RVFL (edRVFL), treat all training samples uniformly, which limits their robustness and effectiveness when applied to real-world datasets containing noise and outliers. Furthermore, the propagation of contaminated features across hidden layers negatively influences the decision-making capability of these models. To overcome these limitations, we propose intuitionistic fuzzy dRVFL (IF-dRVFL) and intuitionistic fuzzy edRVFL (IF-edRVFL) frameworks that enhance model robustness. The proposed models unify intuitionistic fuzzy theory to exploit sample neighborhood information in the kernel space by jointly considering membership and non-membership degrees for each sample. Membership degrees are computed based on the distance of samples from their respective class centroids, while non-membership degrees quantify sample heterogeneity within local neighborhoods. These measures are employed to assign adaptive weights to training samples, enabling effective discrimination among clean, noisy, and outlier data points. Extensive experiments conducted on UCI and KEEL benchmark datasets, with and without the presence of Gaussian noise, demonstrate the superiority of the proposed IF-dRVFL and IF-edRVFL models over existing SOTA fuzzy and non-fuzzy approaches. The source code is available at https://github.com/mtanveer1/IF-edRVFL.

Artificial Intelligence (AI) and Machine Learning (ML) have become powerful tools for supporting and automating complex human tasks. Despite their benefits, growing attention has been directed toward their environmental implications, primarily due to their high energy demands and associated carbon emissions. This concern is particularly relevant in light of the increasing deployment of large-scale models, especially Deep Learning (DL) architectures, which provide advanced predictive capabilities but require substantial computational resources. This paper presents a systematic review of research on Green AI, Green DL, and optimization techniques aimed at reducing the environmental impact of AI models. In addition, we examine and compare several carbon measurement tools for estimating emissions generated by AI algorithms. To complement the review, we conducted an empirical evaluation using a CPU-based experimental setup, in which six DL models were implemented for a multi-label classification task. The objective was to quantify and compare their overall carbon emissions and to determine which stages of the DL lifecycle contribute most significantly to the total footprint. The results show that the training phase is the primary source of emissions. Moreover, the findings reveal that increased architectural complexity does not systematically translate into proportional accuracy gains, highlighting the importance of carefully balancing predictive performance and environmental cost. These results reinforce the need to integrate sustainability considerations into model selection and AI system design.

Transformers have had a profound impact on the world of language processing and computer vision. As efforts to answer the million-dollar question of “How does a Transformer learn?" have been increasing, existing interpretability studies primarily analyze representations at isolated layers or the network as a whole, while the developmental evolution of individual representations and its manifolds across transformer layers remains underexplored. With this work, we aim at providing a comprehensive analysis of the evolution of representations as the representation point cloud transforms across the layers; thereby attempting to isolate layers or establish a trend which comes closer to justifying how and when raw input representations evolve into task-relevant feature representations. Thus, Transformer Geometry Observatory-TGO-IV introduces a topological framework for analysing the evolution of Transformer representations through the lens of Persistent Homology. Rather than studying local geometric properties alone, TGO-IV constructs Vietoris–Rips simplicial complexes from token-level representation point clouds and investigates the evolution of their persistent topological signatures across Transformer layers. The proposed framework comprises complementary topological observatories including Persistence Diagrams, Barcode Diagrams, Betti Curves, Persistence Landscapes, Bottleneck Distance, and Wasserstein Distance, enabling a comprehensive analysis of how the global topology of representation point clouds develops throughout the forward pass.

Recent advances in machine learning have greatly improved breast cancer detection, enabling more accurate and timely diagnosis. Deep learning (DL) models show strong potential for medical image analysis; however, as their architectural complexity increases, their environmental impacts are becoming a growing concern. In this paper, we present a comparative analysis of seven DL models for breast cancer detection on two medical datasets: Breast Ultrasound and BreakHis 400X. The evaluated architectures range from Convolutional Neural Networks (CNNs) and transformers to hybrid models. In addition to performance metrics, we assess CO2 emissions during both training and inference. Our results show that EfficientNet and ResNet consistently deliver strong performance, although with higher CO2 emissions. The selected transformers, such as DeiT-Tiny, perform competitively on both datasets, whereas DenseNet121 achieves lower accuracy. On the Breast Ultrasound Dataset, DeiT provides the most favourable balance between accuracy and energy consumption, whereas on the BreakHis dataset, the ViT and Swin models achieve the best results. Overall, our findings indicate that no single architecture category from the evaluated ones consistently dominates across the two selected datasets. Our results highlight the importance of jointly considering performance, emissions, and dataset characteristics when selecting models for medical applications.

Controllable generation guided by external knowledge is a key requirement in modern generative deep learning applications, enabling the synthesis of samples with explicit constraints on semantic content, structural properties, and variability. In 3D Computed Tomography (CT), such control is essential for clinical applications, including data augmentation, privacy-preserving data sharing, and the simulation of specific anatomical or pathological scenarios. While research on conditional 3D CT generation has expanded rapidly, the diversity of existing approaches makes systematic comparison difficult and obscures fundamental design choices. In this survey, we propose a conditioning-centric taxonomy that organizes the literature along three orthogonal dimensions: the type of external knowledge (K), the knowledge integration paradigm (I), and the generative architecture (A). This factorization defines an explicit design space (K x I x A) that provides a unified perspective on prior work. Using this framework, we systematize existing methods, identify dominant trends and recurring design patterns, and highlight underexplored regions of the design space that point toward promising directions for future research.

Quaternion-valued neural networks and variational quantum circuits (VQCs) both derive local transformations from $\mathrm{SU}(2)$ geometry, yet their performance on classical supervised learning remains poorly understood. We compare real-valued, quaternion-valued, and quantum classification heads on identical frozen features across MNIST, FashionMNIST, and CIFAR-10. CIFAR-10 uses a learned 16-dimensional bottleneck and frozen ImageNet-pretrained ResNet18 features to separate architecture from representation quality. Quaternion classifiers match or approach real-valued baselines while outperforming shallow VQCs. On MNIST and FashionMNIST, quaternion networks nearly equal real-valued MLPs, whereas product-state VQCs show lower accuracy and higher cost. On CIFAR-10, quaternion networks retain 94–97% of real-valued performance and remain stable under a 32-fold increase in dimensionality. Product-state circuits underperform quaternion classifiers, while entanglement gives modest grayscale gains but reverses under pretrained CNN features (9.25 pp degradation vs.\ product-state). Fubini–Study/QFI natural gradients improve geometric alignment but not short-horizon loss reduction vs.\ Adam. A Friedman test on five-seed MNIST detects model differences ($\chi^2=12.796$, $p=0.0051$, $n=5$), with Wilcoxon tests yielding large effect sizes ($d>5$) for QuatNet vs.\ quantum comparisons. For FashionMNIST and CIFAR-10, large effects ($d>2.0$) are the primary statistic given $n=3$. These results indicate that quaternion networks provide efficient, stable $\mathrm{SU}(2)$ alternatives to shallow VQCs on tasks lacking intrinsic quantum structure. Shared local $\mathrm{SU}(2)$ geometry and shallow entanglement are insufficient, within the regime studied, to confer practical quantum advantage. Conclusions are limited to shallow, measurement-limited circuits on such tasks.

Recent advances in generative recommendation (GR) leverage large language models (LLMs) as recommender backbones, enabling LLMs to directly generate recommendations conditioned on item-interaction histories. In these systems, items are often represented through semantic IDs (SIDs) added to the LLM vocabulary as special tokens. Ideally, SIDs imbue item token representations with semantic priors, thereby improving model generalization. However, standard vocabulary expansion typically initializes these tokens as random Gaussian vectors, discarding the SIDs' underlying continuous geometry and forcing the LLM to relearn token relationships from interaction data. To demonstrate the consequences of this design, we first show that training from this initialization tends to organize SID embeddings around item popularity rather than semantics. We further show that, despite partially reducing the reliance on popularity and improving cold item performance, the computationally expensive process of continual pretraining (CPT) fails to reliably recover the original semantic geometry. To address these findings, we propose a simple, parameter-free intervention that initializes SID token embeddings directly from their corresponding centroids in the semantic embedding space. Requiring only a few lines of code and no additional training or inference overhead, this drop-in approach improves pure-SFT Recall@5 by up to 16%, reaches peak performance with up to 40% fewer SFT steps, and improves cold-item Recall@5 by up to 60%. Moreover, on datasets that benefit from additional CPT, centroid initialization reaches comparable performance while requiring half as many CPT epochs. Together, our findings show that preserving SID geometry, beyond shared-prefix structure, provides a simple and effective semantic prior for LLM-based GR.

Mixture-of-Experts (MoE) language models deliver high capacity at low per-token compute, but deploying them cheaply requires compressing their many expert weight matrices. Expert pruning (e.g., REAP) and merging reduce cost but sacrifice accuracy and require retraining the router; low-rank delta decomposition of experts (e.g., D^2-MoE) preserves all experts and the router, but degrades sharply as the expert count grows because a single shared component cannot approximate many near-orthogonal experts. Because MoE expert weights are near-orthogonal, a single shared component (as in prior delta decomposition) scales poorly with the expert count; we show that experts nonetheless organize into functional co-activation communities that are decoupled from weight similarity. Building on this, we introduce LorExperts, a router-preserving compression method that clusters experts, keeps one full-precision dominant per cluster, and represents the remaining members as low-rank corrections to their local dominant. LorExperts retains all experts and the original router (no router retraining). At ~50% expert compression on Qwen3-30B-A3B and Gemma-4-26B-A4B, LorExperts preserves downstream accuracy and perplexity better than the baselines on most of the tasks; the margin over D^2-MoE grows with expert count E. We further give a reconstruction fine-tuning procedure for LorExperts, and BTExperts, a tree organization of dominants and corrections that enables inference-time amortization of shared computation.