Finding high-performing formulations is difficult when few have been measured and experimental budgets are limited. We propose KolTa, an active learning framework that selects batches to discover the top 5% of formulations at a target temperature under a fixed query budget. KolTa combines a three-stage predictor of component contributions, role-conditioned interactions, and anchored temperature response with task-specific soft physics constraints. Its Physics-Residual-Calibrated Upper Confidence Bound (PRC-UCB) acquisition rule combines predicted utility, ensemble disagreement, and a target-temperature physics residual, with nonnegative exploration weights fitted to out-of-fold signed errors. We evaluate the framework on three measured tasks and one simulated task covering binary-mixture viscosity, drug solubility, ionic-liquid viscosity, and electrolyte conductivity. KolTa ranks first or second in mean area under the learning curve (AULC) on every task and outperforms seven of the eight comparator strategies in paired tests, winning 75 of 78 paired comparisons across tasks and seeds against those seven strategies. Across the three measured tasks, its largest gap to the best mean AULC is 0.032, compared with at least 0.056 for every comparator. On Drug, reaching 80% average recovery of the remaining targets across drugs requires an estimated median of 348 additional formulation queries, about 18% fewer than the fastest comparator. Researchers can adapt the framework to diverse formulation problems through component roles and physical constraints.
Multivariate probabilistic forecasting requires modeling cross-variable dependence alongside marginal uncertainty. The challenge is to learn this dependence while controlling marginal noise scales and avoiding dense covariance costs. We propose Factor-Correlated Diffusion (FaCoDiff), which couples a compact conditional statistical estimator with structured denoising. History-conditioned factor loadings use distinct coefficients for forecast-center correction, scale adjustment and noise dependence. Diagonal compensation preserves the estimated marginal noise variances, while low-rank operators enable residual Gaussian likelihood evaluation and noise transport without dense covariance storage or factorization. Factor-derived feature modulation explicitly conditions the denoiser, and a geometry-aware objective accounts for errors after noise transport. The same conditional geometry governs forward perturbations and reverse sampling. Across nine datasets, FaCoDiff achieves the lowest mean variogram score on eight datasets and ranks among the top two on 43 of 54 dataset-metric combinations in the reported comparisons. Matched ablations on Solar and Weather show that factor covariance improves mean Energy and Variogram scores under all four combinations of center and scale correction. Disrupting cross-variable sample pairings while preserving marginal trajectories worsens joint and aggregate scores on both datasets, demonstrating the predictive value of generated cross-variable dependence.
RL for LLM reasoning often plateaus early as its policy sharpens. Policy entropy is studied as both a diagnostic and a control signal, but increasing it alone does not reliably change the eventual plateau in our experiments. We analyze RL's implicit target policy using a sharpness factor $\kappa$ and effective counterpart $\kappa_{\mathrm{eff}}$ that describe how reward or advantage gaps translate into probability-ratio gaps. We introduce reward budgeting, a plug-in advantage construction assigning a fixed total reward budget to each prompt. For binary rewards and independent on-policy rollouts, its expected unclipped score-function update is the exact gradient of a finite-$G$ concave objective with diminishing returns; its advantage gap decreases as more rollouts become correct. For general rewards, a shifted log-mean interpretation holds asymptotically under the stated regularity conditions. On Qwen3-8B and Llama3.2-3B-Instruct, reward budgeting consistently improves RL baselines, with gains up to +12.50 on AMC23 and +10.00 on AIME24 and AIME25. These results support an update-dynamics view of premature convergence and substantial gains without explicitly preserving entropy. Code is available at https://github.com/Qwen-Applications/RewardBudget.
Support vector machines (SVMs) are fundamentally grounded in geometric margin maximisation and currently lack a mainstream probabilistic interpretation. This keeps them theoretically isolated from probabilistic machine learning models that are naturally comparable and combinable. In this paper, we bridge this gap by proving that the dual objectives of both support vector classification and support vector regression are exact, constrained Kullback-Leibler (KL) divergence minimisation tasks between two Gaussian measures. We further show that this same divergence formulation recovers the exact objective function for kernel ridge regression when domain constraints are removed and $L^2$ regularisation is applied. By demonstrating that SVM duals directly minimise relative entropy, our results provide a unified information-theoretic model for standard kernel machines and lay the foundation for unifying broader kernel-based algorithms under probability theory.
Thyroid nodules with indeterminate cytology (Bethesda III) carry an intermediate risk of malignancy, and molecular testing is widely used to decide between surgery and surveillance, but it is costly and not universally available. We investigated whether a convolutional neural network (CNN) applied to routine ultrasound images can stratify Bethesda III nodules, and how it compares and combines with molecular testing. We retrospectively identified 466 nodule records from 418 patients who underwent molecular testing (Afirma or ThyroSeq). A CNN pre-trained on determinate nodules was adapted to indeterminate nodules with soft labels and transfer learning. After removal of duplicate records, 119 nodules with surgical pathology were used for development and 47 were held out for testing. On the test set, the deep learning (DL) model achieved an AUC of 0.73 (95% CI 0.51-0.94), compared with 0.55 (0.38-0.71) for molecular testing (P = 0.15). At a cut-off inherited from the original algorithm, the DL model was more specific than molecular testing (0.95 vs 0.32; P < 0.001), with a higher positive predictive value (0.67 vs 0.21), a lower sensitivity (0.44 vs 0.78; P = 0.25) and a similar negative predictive value (0.88 vs 0.86). All six DL-positive test nodules were molecular-positive, and the DL output split the molecular-positive nodules into a high-risk (4/6 malignant) and a lower-risk group (3/27 malignant; P = 0.01). Surgical verification was strongly driven by the molecular result (99% of positive vs 19% of negative results), and molecular-test accuracy depended on the platform. A logistic regression of the DL output, nodule size and the molecular result dropped the DL term and did not improve on the DL output alone. Ultrasound DL complemented molecular testing for Bethesda III nodules, in particular by identifying molecular-positive nodules at lower risk of malignancy.
Language-subspace interventions aim to improve multilingual reasoning by suppressing language-related components in mid-layer activations. Using MATH and MGSM across four open-weight LLMs, we ask three questions: do estimated language subspaces capture translation-associated variation? Does projecting them out remove language information? And do they improve accuracy without compromising response language? For the first question, we build subspaces using the existing Mean-SVD construction and our DiffPCA. Both capture substantial translation-associated variation and show similar patterns across model depth. For the second, we project out each subspace and train linear probes to classify language from the projected activations. The probes still identify language almost perfectly, showing that neither subspace exhaustively captures language information. For the third, we evaluate several intervention configurations on Qwen2.5-7B, by varying subspace creation methods and intervention operators. We find that the intervention that most improves MGSM accuracy sacrifices response-language fidelity, with only one intervention improving joint accuracy$\unicode{x2014}$answers that are correct and in the target language$\unicode{x2014}$slightly, by 1.60 points. Interestingly, subtle changes in the prompt alone can reproduce the accuracy gains albeit with significant damage to language fidelity. We argue that causal interventions must be judged on more than accuracy, to fully account for language substitution trade-offs.
Despite decades of progress, image codecs remain difficult to adapt to diverse real-world compression scenarios for two main reasons. First, most image codecs are optimized for a specific objective and require substantial redevelopment when the compression goal changes. Second, they typically assume clean inputs, whereas sensor-captured images often contain degradations that alter rate-quality behavior. To address these challenges, we propose Practical Image Coding Agents (PICA), a multi-agent framework that adapts existing codecs to diverse user requirements and potentially degraded inputs. PICA consists of four collaborative agents: a Perception agent that interprets user intent and analyzes image content, an Improvement agent that jointly considers bitrate and reconstruction quality to guide pre-coding restoration, a Coding agent that refines quality-control settings from global to regional levels, and an After-coding improvement agent that addresses residual degradations and compression-induced artifacts through restoration guided solely by image quality. To enable codec-agnostic regional adaptation, we introduce Agentic ROI Coding (AROIC), which treats the codec as a black box and allows the agent to automatically schedule, execute, evaluate, and reflect ROI coding plans without architectural modifications or codec-specific retraining. We integrate these four agents with a Scenario-Guided Quality-Rate (SGQR) evaluation strategy to enable more effective scenario-adaptive compression across heterogeneous codecs. Comprehensive experiments on clean, noisy, and blurred images across standard, perceptual, and machine-vision compression scenarios demonstrate that PICA consistently enhances the compression performance of both the conventional H.266/VVC codec and the neural Cheng2020 codec, outperforming competing compression methods on the Kodak, CLIC2020, and HEVC datasets.
We study the minimization of the average of $n$ convex $(L_0,L_1)$-smooth functions: the Hessian norm of each function is bounded by an affine function of its gradient norm. We assume known upper bounds $R$ on the distance from the starting point to a minimizer and $\delta$ on the difference between the minimum of the average and the average of the infima of the components. We derive upper and lower bounds on the oracle complexity of reaching accuracy $\varepsilon$ in function value in expectation. Up to logarithmic factors, both bounds are of order $n(1+(L_1R)^{2/3}) +\sqrt{\frac{n(L_0+L_1^2\delta)R^2}{\varepsilon}+n(L_1R)^2}$. The upper bound is attained by the proposed method YETI. The method combines ideas of accelerated proximal methods and variance reduction methods with a heuristic for choosing additional points. The number of oracle calls made by the heuristic is bounded, which preserves the upper bound on the total number of queries needed to reach the target accuracy. In the experiments, YETI is compared with first-order methods on synthetic problems and Poisson regression problems. We also examine the stopping criterion and the effect of the heuristic separately. The lower bound is obtained by an explicit construction of families of instances that are hard for any first-order method.
Selecting small, effective instruction-tuning datasets requires identifying which aspects of the original data to retain. We introduce Follow the Features (FtF), a data-selection method that uses sparse autoencoder (SAE) features to characterize training data and guide subset construction. FtF ranks features using dataset-level activation statistics, scores samples by their expression of highly ranked features, and incorporates feature resampling to encourage diversity. The LLM-SAE pair used for selection is independent of the model being fine-tuned, enabling reuse across target models. We evaluate FtF on two model-dataset settings at a training budget of 100 samples with a fixed sequence length of 256 tokens. FtF achieves performance competitive with GRAPE and S2L, with relative rankings varying across settings. On Qwen 3 8B trained on Alpaca, FtF exceeds full-dataset training by 1.59 percentage points in the harmonic mean of an AlpacaEval-based win rate and Open LLM Leaderboard performance, although full-dataset training retains a higher Leaderboard score. These results support SAE feature statistics as a useful signal for data-efficient instruction tuning.
Knowledge distillation is a technique used to transfer the expertise of large teacher models into small, efficient student models. Recent advances have integrated the distillation of internal activations, more directly training students to learn similar representations as the teacher. One of the core challenges of internal distillation is the dimensional incompatibility between the student and teacher internal representations. Existing methods add learnable matrices to project the larger teacher activations into the smaller representation space of the student. This is a notable point of inefficiency, as learning additional parameters for the projectors increases training costs, only for them to be discarded immediately after training. We introduce RAPID (RAndomized Projection for Internal Distillation) a novel distillation methodology which eliminates the wasteful projection learning by leveraging randomized methods. RAPID utilizes an efficient transform construction based on the Johnson-Lindenstrauss lemma to compress the teacher representations into the student embedding space with strong theoretical guarantees of structural preservation. RAPID also employs a combination of exponential moving average (EMA) and homoscedastic uncertainty weighting in a novel multi-objective function designed for internal distillation. Our experiments in post-training for reasoning demonstrate that RAPID increases training efficiency by as much as 21.1% while maintaining performance comparable to existing methods.
Diabetic macular edema (DME) is a leading cause of severe vision loss among working-age adults. Although intravitreal anti-vascular endothelial growth factor (anti-VEGF) injections currently serve as the first-line therapy for DME, a notable subset of patients fails to respond despite the substantial cost and clinical burden, suggesting they may benefit more from alternative interventions. Automated, noninvasive pre-treatment response assessment could therefore spare non-responders from the current trial-and-error treatment paradigm. In this study, we propose a novel conditional adversarial diffusion framework that synthesizes post-treatment optical coherence tomography (OCT) images from baseline scans, enabling the derivation of each patient's initial response to anti-VEGF therapy. The model was developed with subject-level 5-fold cross-validation on a US cohort (n = 127) and externally validated on an independent Iranian cohort (n = 44). Benchmarked against fourteen baseline and state-of-the-art translation models, the proposed model achieved the best treatment-response prediction performance among the methods evaluated for that task on both cohorts, with balanced accuracies of 85.7% and 93.0% on the US cohort and 85.2% and 95.3% on the external Iranian cohort at the 10% and 20% central subfield thickness reduction thresholds, respectively. Ultimately, this framework offers a decision-support tool to help ophthalmologists anticipate initial anti-VEGF outcomes in DME prior to treatment initiation, toward personalized therapy selection. The source code will be made publicly available upon acceptance.
Understanding how neurons respond to visual stimulus is an important question in both deep learning and neuroscience. It has significant implications for enhancing the interpretability of black-box artificial neural networks and for understanding the visual representation in the human brain. We propose a structured visual representations landscape and design an activation-based prior that allows us to effectively regularize the landscape with activations from brain regions or neural network units. Our model Vis-Lens integrates a variational autoencoder and a diffusion model into an image generation framework. It enables the generation of natural, realistic preferred images by directly modifying the activation-regularized latent codes without per-image iterative optimization. We demonstrate the effectiveness of our framework on both artificial neural networks and human visual cortex using functional magnetic resonance imaging (fMRI) data. Our framework improves activation modulation and image realism compared with a recent state-of-the-art method for generating preferred visual representations of these neural targets. Code is available at https://github.com/NeuroAIHub/Vis-Lens.
Parametric human body models such as the Skinned Multi-Person Linear (SMPL) model provide compact representations of human shape and pose, but their reconstructions remain affected by both errors in estimated latent parameters and limitations in the decoder's expressive capacity. We propose a framework for refining these reconstructions by exploiting the geometric discrepancy between an initial decoded surface and a target mesh, without requiring vertex correspondences between them. The framework comprises three complementary approaches. Latent-VaRefine uses varifold distance gradients to correct the estimated latent code while preserving the original parametric representation. Vertex-VaRefine processes multiscale varifold-gradient features with a DiffusionNet backbone to predict vertex displacements in a canonical pose, yielding an approximately pose-equivariant refinement pipeline. ElaOpt directly optimizes surface deformations under elastic regularization and normal direction constraints. On AMASS, Vertex-VaRefine reduces the mean vertex-to-vertex error of OmniFit reconstructions from 1.181 to 0.342 cm, a 71.0 percent reduction, while combining Latent-VaRefine with ElaOpt reduces the asymmetric Chamfer distance of ArtEq reconstructions from 1.681 to 0.717 cm, a 57.3 percent reduction. Evaluations with SMPL-H and SMPL-X demonstrate the framework's adaptability to different body models, while additional experiments on remeshed targets and raw scans assess its robustness to changes in discretization and geometry beyond the parametric model. These results highlight the complementary strengths of latent space correction, learned vertex refinement, and elastic optimization, offering different trade-offs between preserving the model parameterization and improving geometric fidelity.
Recent advancements in sign language processing can be attributed to developments in both large-scale dataset collection and algorithmic complexity in areas such as pose estimation techniques, signal processing through Transformer models, and model optimization. However, only a handful of sign languages, such as the American, British, and German, from over 200 globally recognized sign languages, reap the benefits of such advancements, while the rest lag behind. In this study, we aim to address this technological disparity by proposing a framework for the automatic collection and preparation of sign language datasets, with very minimal human-labour intervention, using the South African Sign Language (SASL) as a testbed. Specifically, we automatically collect SASL video content using public resources such as news broadcasts. Additionally, we convert their accompanying audio streams into text using available Automatic Speech Recognition (ASR) tools. Furthermore, we crop the sign language area in the videos and use gradient descent to locate valleys of non-motion to determine end-of-sentence points within long videos to create our weakly sentence-aligned parallel corpora -- the Maswao (Signs) dataset. Finally, we benchmark the weakly aligned dataset on the task of Sign-to-text Translation using Recurrent Neural Networks and Transformer models to compare their individual intricacies. Our results show that Transformer models proposed in the literature outperform recurrent baselines, with the best-performing Transformer achieving BLEU-1 to BLEU-4 scores of 13.78, 5.21, 2.52, and 1.54, respectively, as well as ChrF = 18.09 and ROUGE-L = 10.51. These results indicate that the task remains promising and warrants further investigations.
Interleaved multimodal chain-of-thought reasoning improves the visual reasoning capabilities of multimodal large language models by dynamically incorporating visual evidence during reasoning. However, existing methods primarily select evidence from historical context. When the reasoning focus shifts, history-guided evidence may remain focused on prior regions and fail to cover the visual information required by the next step. We refer to this lag between visual evidence and subsequent reasoning demands as Temporal Evidence Myopia. To address this issue, we propose Lookahead Prediction-guided Adaptive Visual Refocusing (LPAVR), a training-free test-time framework. LPAVR transforms Multi-Token Prediction (MTP) from a decoding acceleration mechanism into a lookahead signal for anticipating upcoming reasoning intent. During MTP decoding, the draft head generates multiple future tokens, while the target model verifies them in parallel and retains the longest prefix consistent with its decoding trajectory. This target-verified prefix contains near-future semantics, which LPAVR uses to construct lookahead-guided visual evidence and assess whether history-guided evidence covers the indicated regions. If the coverage is sufficient, the history-guided evidence is retained; otherwise, LPAVR switches to the lookahead-guided evidence. This better aligns the visual context with upcoming reasoning and improves consistency between draft predictions and target-model decoding. Experiments on multiple multimodal reasoning benchmarks show that LPAVR improves both reasoning accuracy and the acceptance rates of MTP speculative tokens, jointly enhancing visual reasoning and speculative decoding efficiency.
Large language models (LLMs) can generate visual assets through text-based formats such as SVG and CAD, but these symbolic representations describe images using high-level primitives rather than explicit raster values. We ask whether LLMs can instead generate pixel images directly through their existing vocabulary, without a learned image tokenizer or decoder. We introduce PixelBench, a benchmark for text-to-pixel (T2P) and image-to-pixel (I2P) generation across multiple resolutions, together with a complementary pixel-understanding evaluation. Models generate an RGB color palette followed by a grid of indices specifying the palette color of each output pixel. In T2P, the strongest models produce recognizable images directly from text. In I2P, the strongest models outperform palette-matched downsampling and specialized pixelization algorithms at low resolutions. Pixel-understanding performance correlates strongly with generation performance. A resolution-scaling study shows that pixel-generation performance often improves initially before declining at high resolutions. Finally, we introduce Pixel Reconstruction Tuning (PRT), which augments visual instruction tuning with caption-palette-grid trajectories. We post-train Molmo2-4B from its pretraining checkpoint by mixing general instruction data with our curated PRT data, improving performance across 16 visual-centric benchmarks, indicating that explicit pixel reconstruction can provide useful auxiliary supervision.
Search-based planning remains a powerful policy improvement paradigm in reinforcement learning, driving milestone successes in discrete domains. However, extending this paradigm to continuous control exposes a structural mismatch: search evaluates a finite set of action candidates, whereas continuous RL typically relies on Gaussian policies. We show that winner-action distillation into a continuous Gaussian induces a \textit{projection error} and a stepwise variance-expanding gradient tendency when the selected action is sufficiently far from the current mean. Empirically, this tendency is associated with elevated terminal variance and a return gap on the evaluated tasks. To address this, we propose \textbf{Iterative Gumbel Planning (IGP)}. By adopting a per-dimension factorized categorical policy, IGP naturally aligns the policy support with the search candidates, eliminating the projection error and reducing the residual approximation to a bounded discretization error. To bypass the combinatorial explosion of high-dimensional discrete action spaces, IGP pairs Gumbel Top-$K$ candidate sampling with an MPPI-style multi-round iterative search to refine the policy at each state. Empirically, IGP achieves state-of-the-art aggregate performance and sample efficiency on DMControl and HumanoidBench. Furthermore, it produces more stable action sequences with lower temporal variance, demonstrating its practical potential for real-world robotic deployment.
Large Language Models (LLMs) are increasingly deployed in multi-agent settings, exhibiting emergent collective behaviors, yet the role of curiosity as a driver of such dynamics remains understudied. We introduce CuriousTogether, a framework evaluating curiosity-driven multi-agent LLM interactions across three complementary dimensions: novelty-seeking, epistemic and social curiosity. Across eight tasks spanning reasoning, creative generation, and social bias mitigation, curiosity-guided multi-agent settings yield improvements over non-curious ones on seven out of eight tasks, with gains of up to 20.6% in story generation (semantic diversity) and a 14.9 pp reduction in stereotype rates. Furthermore, country-conditioned cultural grounding show benefits for story generation and social tasks, but can hurt precision-oriented reasoning. Finally, we show that inference-time multi-agent curiosity-based approach offers favorable quality-token trade-offs, compared with single-agent baselines. Together, these findings support curiosity-guided interaction as a practical mechanism for improving multi-agent collaboration, with downstream benefits for reasoning, creative, and societal tasks.
The absolute trajectory error (ATE) is the standard measure of camera-motion accuracy, from SLAM to feed-forward 3D reconstruction. On the common Sintel protocol, an image-free constant-velocity line has a lower mean ATE than 10 of 11 recent models re-run with official code and than 85 of 110 published methods. On Lightspeed it beats 5 of 9 models and 10 of 11 published methods, while on TUM, ScanNet, DyCheck and ADT every model beats it. Two properties of ATE's alignment explain this. First, it maps any moving constant-velocity prediction onto the line that best fits the true path, so the line scores the path's regression residual. A residual signal-to-noise ratio (SNR) criterion tells when moving a prediction towards its constant-velocity fit lowers ATE. With the alignment's first-order refit, it accounts for the outcome of all but one of 4010 model-clip evaluations, including all 504 on three benchmarks registered before any model ran. The true path's straightness flags Sintel and Lightspeed. Second, the alignment ignores where cameras look: rotating a trajectory by 30 degrees against its orientations leaves ATE unchanged yet worsens the static-point reprojection of every model on Sintel, TUM and ScanNet, and the line reprojects worst on Sintel by geometric mean. Fitting the alignment's rotation to the predicted orientations gives pose-anchored ATE (PA-ATE), a one-function change at about ATE's cost (37 vs. 34 $\mu$s per clip, and 61 vs. 66 $\mu$s with the rotation error). It is never below ATE and, as we prove, equals the translational relative pose error over all frame pairs divided by $\sqrt{2}$ when the orientations agree. It ranks every image-free reference below every model on Sintel, follows the reprojection error where ATE does not, and largely preserves the models' order. Its rotation error also exposes Lightspeed ground-truth orientations that the images contradict.
We study how the layers in deep neural networks (NNs) converge, using Wilsonian Renormalization-Group (RG) ideas to organize the spectral picture. During training, these layers often develop heavy-tailed power-law (PL) spectra, and the strongest checkpoints often appear near the PL exponent $\alpha\approx 2$. This pattern raises a simple question: why should this value mark a well-trained layer? The answer is both statistical and physical. Our statistical hypothesis is that $\alpha=2$ marks the boundary between stable if not optimal learning and overfitting via non-self-averaging. The physical answer is that the same value makes the spectral energy gain scale evenly across logarithmic spectral scales, which is the proposed Wilsonian marginal condition. We then apply the ERG trace-log condition, the proposed SETOL theory, which we justify by taking volume-preserving, scale-invariant transformations to remove arbitrary layer scale along the proposed RG flow. The resulting spectral RG flow connects the HTSR and SETOL theories to optimizers such as Muon, and motivates a new optimizer extension, WW-PGD. Together, these results point toward a spectral Renormalization-Group view of learning, in which NNs may approach shared spectral structure in the settings studied here.