Low-bit quantization reduces the inference cost of LLM agents but can compromise task success. Examining failed rollouts, we find that quantized agents overlook task requirements or environmental conditions, producing environment transitions that diverge from successful full-precision references. Correcting these behavioral divergences remains difficult: offline data may miss the student's failure contexts, while its own rollouts may provide few correct responses in those contexts. We introduce SteerDistill, a quantization-aware distillation method in which a full-precision teacher guides the student to revise failed responses and iteratively refines its guidance using feedback from student execution. This closed-loop teacher steering produces locally verified behavioral corrections. We internalize these corrections by training the student without guidance to match the steered student's predictions at corrected turns and the full-precision teacher's at other turns, enabling unassisted inference. Across five agentic benchmarks and three models, SteerDistill exceeds the strongest quantized baseline by 1.67--2.18 points in average score under W4A16 and stays within one point of BF16. On $\tau^2$-Bench, it outperforms QeRL by 1.84 percentage points in success rate using 67% fewer training GPU hours and produces shorter successful trajectories than the quantized baselines.
https://arxiv.org/abs/2611.13410
Electronic health record (EHR) foundation models trained on longitudinal patient trajectories have demonstrated strong performance across diverse clinical prediction tasks. However, their clinical reasoning capabilities remain constrained by next-token prediction on limited and incomplete EHR data. To address this, we propose a reinforcement learning (RL) fine-tuning framework that treats EHR foundation models as generative policies over patient trajectories. We formulate common clinical prediction problems (e.g., hospital readmission) as event-conditioned, time-windowed reasoning tasks. We then design time-aware, rollout-sensitive rewards to account for finite rollout lengths and temporally inconclusive outcomes. We find that RL fine-tuning consistently improves over pre-trained backbones and strong baselines. Notably, it enables smaller models to match larger pre-trained models in data-limited regimes and induces positive transfer across tasks. Further analysis shows that RL fine-tuned models generate trajectories with stronger structural and semantic alignment to ground truth and greater downstream utility.
https://arxiv.org/abs/2609.12277
We present a neuro-symbolic framework that pairs online-adaptive Fuzzy Cognitive Maps (FCMs) with long short-term memory (LSTM) networks for financial forecasting. Financial markets exhibit regime shifts and nonlinear dynamics. Existing models often trade off one of these key properties: these models capture temporal patterns but hide causal drivers, or they recover causal graphs that cannot adapt post-training. We close this gap with two frameworks. Framework I constructs an FCM from live dual-source price series and macroeconomic indicators, and adapts its edges online through an error-driven Hebbian rule. We introduce the Asymmetric Additive Residual Fusion method to seamlessly combine the FCM and LSTM predictions under Framework I. Framework II constructs a literature-grounded, LLM-augmented FCM where every edge carries a verifiable citation that gives an auditable map of factor interactions. On NASDAQ Composite data, Framework I matched the performance of the LSTM baseline on state-directional accuracy and outperformed LSTM by 7.6 percentage points on the return-sign accuracy. Framework I outperformed standalone FCM models across both evaluation metrics. The adaptive maps converged to new regimes within 10-15 trading days, and Framework II center-node FCMs scored 68.9% to 85.7% on the state-directional accuracy.
https://arxiv.org/abs/2611.13388
Practitioners often combine independently trained models while keeping their weights frozen. These experts are commonly combined through their final predictions, although their internal features may retain useful information that those predictions omit. We ask when combining features improves on combining predictions, and how to make that comparison fair. We introduce a reusable evaluation protocol, apply it to 21 frozen deepfake detectors from DF40 (Yan et al., 2024), and audit how evaluation choices change the conclusions. The protocol uses strong prediction baselines, compares simple and larger feature combiners, varies architecture and training data separately, and reports results by held-out generator with train/test overlap controls. Under leave-one-generator-out evaluation, logistic regression on the features of six architecture-diverse experts improves macro AUROC by 0.081 over the best tested prediction combiner in each fold, with gains on seven of eight generators. The larger network shows no clear improvement over the linear model (approximately 3.5M versus 10K parameters), suggesting that access to richer information matters more than additional combiner capacity on this pool. Features may help when they retain information that the predictions omit and that transfers to new generators. Predictions may be enough when the strongest prediction baseline already performs well, while neither approach reliably compensates for poor expert transfer. Weakening prediction baselines on the overlap-affected standard split inflates the apparent feature advantage by up to 0.172 AUROC. Overlap-aware evaluation also changes the estimated gain, though the contrast does not isolate overlap. Results by generator and expert pool show where the advantage holds, weakens, or reverses. Without these controls, the same benchmark can support different conclusions.
https://arxiv.org/abs/2611.13356
Protein engineering requires navigating rugged fitness landscapes, where small sequence changes can produce large functional differences, but experimental screening can only sample a small fraction of sequence space. Thus, a central challenge is to efficiently traverse the large combinatorial space of mutations in silico. We introduce ProTEUS, a mutation-resolved world model for closed-loop protein evolution. Pretrained on deep mutational scanning assays, ProTEUS learns a joint representation of sequence, evolutionary, and structural information using MutationFormer, a mutation-biased attention architecture. Variants are represented as states in this joint latent space, and mutations define actions whose functional consequences are predicted. To support exploration of many mutants, we developed a fast, physics-constrained all-atom predictor that utilizes explicit geometric constraint enforcement to achieve >10 fold increase in speed than AlphaFold3 with guaranteed physical validity. Lastly, ProTEUS refines its prediction and understanding of protein activity through closed loop experimental feedback to guide evolution toward a desired phenotype. ProTEUS surpasses established models in both zero-shot and supervised fine-tuning mutation prediction tasks. We experimentally demonstrate its utility in two closed-loop multi-objective engineering campaigns on aldehyde dehydrogenase and levansucrase, yielding variants with >10 fold higher activity than wild type. ProTEUS provides a world-model framework that connects virtual evaluation of mutational outcomes with experimental learning for efficient protein evolution.
https://arxiv.org/abs/2611.13348
Accurate localization of abnormalities in three-dimensional (3D) CT scans is essential for linking automated predictions with clinically relevant pathological findings. We propose Uni-Loc, a vision-language framework that unifies pathology classification and 3D localization using scan-level supervision. Uni-Loc derives both predictions and localization maps from spatial similarities between pathology-conditioned queries and image features. A token-to-voxel upsampling module enables high-resolution localization, while gradient-guided thresholding identifies pathology regions without additional learned parameters. Evaluated on five localization tasks comprising 355 annotated CT scans, Uni-Loc achieves macro-averaged voxel-level AUROC/AUPR of 96.1\%/34.5\%, outperforming relevance propagation (80.7%/13.9%) and gradient saliency (88.7%/8.1%). Uni-Loc also achieves classification AUROC of 88.39% on CT-RATE and 78.03% on RAD-ChestCT, improving over a matched baseline. These results demonstrate effective 3D pathology localization without requiring pathology-specific voxel-level supervision.
https://arxiv.org/abs/2611.13330
Neural dynamics are a central feature of biological neural systems. In the visual cortex, the functional role of neural dynamics is poorly understood, particularly for static stimuli. The statistics of still natural images have long been used by models that successfully predict firing in early visual cortex; those models assume that neurons represent inferences about latent image features, but often ignore the dynamics of said inference. Here, we test if inferential dynamics in images explain visual-cortical dynamics. To do so, we present a biologically plausible model of cortical computation which assumes that the inference of features is coupled with inference about how those features are organized into meaningful parts, a function also known as segmentation. Given an input image, our model iterates between inferring the features and the segments, through recurrence and feedback, thus inducing dynamics in the neural representation of features. For these neural dynamics, we derive theoretical and exact predictions. Our model predicts both gain and variability in evoked responses to natural images, and we show that it captures classical single-neuron experimental findings such as gain dynamics and variability decay during stimulus presentation. The model also predicts that those dynamics are organized around inferred segment structure and more heterogeneous than previously thought, reflecting the complexity of inference for natural scenes. Finally, our analytical formulation provides a clear interpretation of our results. Insights from our normative model could generalize to segmentation problems in other experimental modalities, and help address the cost of inference in artificial networks with iterative computation.
https://arxiv.org/abs/2611.13316
Temporal Sentence Grounding (TSG) aims to localize temporal segments in untrimmed videos that correspond to natural language queries, a critical task for video understanding with applications in video retrieval, summarization, and human-computer interaction. Despite significant progress driven by transformer-based models, existing TSG methods suffer from two key limitations: quantization errors due to temporal downsampling, which misalign segment boundaries, and suboptimal ranking of predictions based on semantic matching rather than localization precision. To address these challenges, we propose the Adaptive Boundary Refinement and Ranking (ABRR) framework, a novel, model-agnostic post-processing approach that enhances the performance of any TSG model without retraining. ABRR integrates three innovative components: (1) Temporal Distribution Smoothing, which mitigates noise in predicted score distributions using Gaussian kernel convolution; (2) Boundary Refinement via Optimal Transport, which leverages the 2-Wasserstein distance and Sinkhorn iterations to align coarse segment boundaries with true temporal moments, grounded in Monge-Kantorovich optimal transport theory; and (3) Quality-Driven Re-Ranking, which reorders predictions using a composite score balancing localization precision and semantic relevance. By correcting quantization errors and prioritizing accurately localized segments, ABRR achieves substantial improvements over state-of-the-art TSG models. Extensive experiments on benchmark datasets, including Charades-STA, ActivityNet Captions, and QVHighlights, demonstrate consistent gains at stringent IoU thresholds (e.g., IoU=0.7), alongside enhanced ranking performance. The framework's computational efficiency and compatibility with existing models make it a versatile and impactful solution for real-world TSG applications.
https://arxiv.org/abs/2611.13310
We introduce Latent Barrier Steering (LBS), a hierarchical semantic-to-path framework for safe flow-matching-based generative planning. Existing control-barrier-function (CBF) safety filters for diffusion and flow-matching planners act locally on generated samples, either during sampling or after prediction. Such repair is effective for small violations, but can struggle when a new constraint blocks the whole selected behavior mode rather than merely perturbing it locally. LBS decouples behavior selection from path certification through two coordinated interventions: it first steers the generator's behavior latent toward rollouts with larger safety margin, then decodes the steered latent and applies a path-space CBF quadratic program (QP) only for residual violations. We show that latent steering increases the margin available to the path-space corrector, yielding finite-time recovery guarantees and reduced correction burden. Experiments across navigation and robot end-effector planning tasks show that LBS preserves safety while improving goal reaching and reducing correction burden compared with path-level safety filters.
https://arxiv.org/abs/2611.13227
We study AI-assisted measurement when one text label applies to many administrative records, the target is a monetary total, and the human readings that anchor the machines disagree. The testbed is 121,070 grant-purpose texts from US tax filings (799,675 records, \$126.4B), labelled by a rule-based classifier and three language models, with a designed sample of 651 texts read independently by three people under one codebook. We state exact conditions for a fixed-predictor difference estimator and for a cross-fitted predictor whose folds are assigned to the whole frame before sampling, under Poisson sampling, and assess the implemented fixed-size procedures by simulation: 8 predictors whose own totals span \$11.2-15.0B agree to within \$0.4B once rectified. Two rules that accept machine labels at a stated dollar-weighted error, weighted conformal p-values under a dollar measure and Learn-then-Test on the design estimate of the error, are evaluated on nested held-out folds and on re-draws of the design with synthetic truths; their error control in deployment is not established, and the count-based rule accepts no text under a dollar-drawn sample. In a replay that charges every human label, a two-phase review policy whose calibrator is fitted on its own first phase reduces the RMSE of the total by 28% at a budget of 100 texts relative to dollar-proportional sampling, while pure Neyman allocation fails with an uncalibrated predictor. Independent readers shift the estimated total by \$3.93B, 2.2 times the within-reader sampling interval, and institution-level intervals under-cover at every size. Every headline result is re-run under each reader. The results identify both the benefits and the limits of using learned predictions to decide what a person should verify.
https://arxiv.org/abs/2611.13215
Machine learning-based resource management (RM) models degrade under out-of-distribution (OOD) inputs at runtime, yet existing approaches in the literature rely on a single global error margin that is overly conservative and fails to distinguish reliablepredictions from catastrophic failures. We propose Lightweight Runtime EXplainability (LEX), the first post-hoc XAI framework for input-specific runtime confidence estimation in ML-based RM. LEX clusters the embeddings of the training input space using a Gaussian Mixture Model and associates per-cluster prediction error bounds to input samples, enabling fine-grained runtime confidence signals without retraining. Evaluated on a state-of-the-art LSTM-based IPS forecasting model on Intel Alder Lake, across 34 benchmark applications, LEX achieves ~94% containment on unseen workloads and produces ~17% tighter per-input confidence bounds than global baselines. With its negligible runtime overhead, LEX advances the field by providing explainable and reliable input-specific confidence signals of ML predictions, thereby making ML-based resource management resilient to OOD at runtime.
https://arxiv.org/abs/2611.13199
We study the problem of cross-sample modality attribution, which involves measuring how one modality in one sample affects the learning outcome in other multimodal samples. This problem is crucial in a multimodal system because a data modality (such as a text description) may be redundant for one sample (a self-explanatory image), but the same data modality would otherwise be uniquely informative for another sample, or informative when combined with other modalities. Prior work focuses on influence estimation at the level of one sample, and it remains an open question how one should capture the interactions across both samples and modalities. To capture scenarios like the above, we propose structured kernel surrogate models, which are surrogate mappings from both samples and modalities to the prediction outcome when they are trained together. We design an algorithm, SKS, which applies kernel ridge regression with a structured, compositional kernel on the selection states of individual samples and modalities. Specifically, the kernel function multiplies kernel-based similarity scores between every pair of data modalities for one sample and a pair of two samples. Additionally, we design a gradient-based estimation algorithm to efficiently estimate the structured kernel surrogate model. We evaluate our approach on two multimodal benchmarks, spanning a range of data modalities including satellite images, road networks, text, audio, and tabular data. We find that SKS achieves $57.8$% higher correlation with the true prediction outcomes than existing methods. In downstream data selection, SKS improves the average accuracy by $1.93$% over baselines. Models with data selected by SKS have lower Hessian traces, indicating a flatter loss surface. Using SKS to select satellite images identifies road segments, where accident scenes commonly occur, as opposed to background images.
https://arxiv.org/abs/2611.13198
Image geolocalization predicts the location at which a photo was taken. Accurate geolocalization requires discriminating between both physically and semantically adjacent areas on Earth. However, prior diffusion-based approaches predict locations by generating coordinates, which are semantically unstable over space. We avoid this limitation by learning geolocalization behavior in the pretrained representation spaces of location encoders, which provide both positional and semantic understanding. To make this space usable for coordinate prediction, we introduce a Location Autoencoder (LAE), which pairs a frozen location encoder with a learned decoder mapping embeddings back to coordinates with kilometer-level accuracy worldwide. In the LAE space, we train DIG, a latent diffusion model that transports noise to location embeddings conditioned on image features. DIG consistently outperforms coordinate-based generative models and sets a new state of the art for parametric geolocalization on three benchmarks: OSV-5M, Im2GPS3k, and YFCC4k. Notably, DIG is the first generative model to outperform retrieval-based methods, and further improves retrieve-and-rerank systems when used as a prior. Code and models will be released.
https://arxiv.org/abs/2611.13180
Graphical user interface (GUI) grounding maps a screenshot and a natural-language instruction to the screen location where an agent should act. Zoomed views can reveal small targets but may lose the context needed to identify them. We introduce GroundCascade, a training-free controller that uses agreement between a grounder's full-screen and zoomed predictions to decide when to consult a complementary model. It selects between zoomed predictions while preserving corroborated full-screen locations. An optional Finalizer predicts from an unmarked joint crop when disagreement remains (Full); omitting it yields the two-stage Selective variant. GroundCascade reaches 85.33% on ScreenSpot-Pro with Indeed-UI-32B/8B/32B and 88.43% on the visible-target OSWorld-G-Refined subset with Qwen3.6-27B/3.8-27B/3.8-27B, gains of 2.66 and 2.35 percentage points over the stronger standalone two-pass grounder. On UI-Vision, GroundCascade reaches 72.20% category-balanced accuracy, with Selective slightly higher at 72.28%. Across these three evaluations, GroundCascade in Full mode uses 2.42-2.73 calls per request, 31-39% fewer than unconditionally acquiring both models' two-pass predictions. At matched second-model call budgets, selective acquisition gains 1.52-2.56 points over standalone Two-pass, compared with 0.22-0.69 points from stratified random acquisition. GroundCascade leads the official ScreenSpot-Pro leaderboard and exceeds reported GUI-specialist comparisons on OSWorld-G-Refined and UI-Vision, subject to differences in external evaluation protocols.
https://arxiv.org/abs/2611.13177
Effective teaching requires reasoning about the learner as well as the task, a capacity that Theory of Mind describes as inferring another agent's latent state from its behavior. We ask whether a stronger LLM teaches a fixed student better when it applies this principle. The student is an LLM agent acting through a harness that fixes its tools, context, and action budget, so a procedure that helps one model can exhaust the budget of another. We introduce Student-Aware Teaching (SAT). The teacher builds a learner model that pairs an empirical profile with hypotheses about the student's limitations, tests these hypotheses with diagnostic probes, and writes reusable skills for the diagnosed failures. A skill enters the library only after it raises success on matched development tasks. With Claude-Sonnet-5 teaching DeepSeek-V4-Flash, SAT reaches 86.46% task success across four primary runs, leading both a profile-free teacher with the same framework (81.04%) and the untaught student (75.00%) in each run. Removing the analytical learner model lowers both prediction and success, and under misleading learner profiles none of the 48 candidate skills passes admission. When DeepSeek-V4-Flash and Gemini-3.8-Flash exchange libraries, each performs better with its own. Saved libraries keep average gains of up to 11.1 points after the teacher leaves. The benefit depends on a student that can follow instructions and a deployment route that delivers them.
https://arxiv.org/abs/2611.13144
Recent work on latent-free diffusion transformers has demonstrated the feasibility of training image generators without a latent VAE. This avoids compression artifacts from the VAE and simplifies the training pipeline. In particular, JiT showed that even a standard transformer can be trained to generate images when the model is trained to predict the final image ($x$-prediction). Even though JiT unlocks training efficiency comparable to latent models, it still suffers from reduced image quality likely arising from modeling the image as a collection of large patches. In this work, we revisit JiT and show that adding simple convolutional pixel-pathway layers in a post-training phase significantly improves generation quality. Moreover, we show that the added pixel-pathway does not need to be active in the early denoising stages, and can be activated as late as in the last $20\%$ of the denoising process. Our experiments on ImageNet@256 show improvements compared to the original JiT results, achieving an FID of $1.78$ (from $2.34$ with corrected sampling), without using additional losses.
https://arxiv.org/abs/2611.13120
In the absence of extrinsic rewards or task supervision, human behavior is guided by intrinsic motivation---neural signals that drive curiosity, creativity, play, and other essential facets of intelligence. Intrinsic motivation plays an important developmental role in building robust world models that enable generalization and adaptation. Yet existing studies leave open how a broad range of intrinsic motivations interacts with world-model representation learning to support downstream control. This limits our understanding of which intrinsically collected experiences build reusable models. To address these gaps, we integrate 24 intrinsic objectives, including established methods and explicit variants spanning curiosity, entropy, and competency, into a common model-based reinforcement learning framework. We evaluate the world models they build through zero-shot, few-shot, and multi-task control in locomotion, manipulation, and open-ended gaming environments, using three representation-learning strategies: observation reconstruction, latent prediction, and redundancy reduction. Curiosity-driven exploration with observation reconstruction yields the strongest aggregate family profile, although the best individual pairing depends on the experiment. In compositional vision-based tasks, intrinsic motivation derived from both episodic and lifelong novelty---providing a complementary memory to the world model's generic recurrent state---achieves the best zero-shot transfer and few-shot adaptation. These findings motivate designing intrinsic motivation and representation learning jointly. Our results demonstrate world models as a powerful tool for measuring the quality of intrinsically motivated experience, and positions intrinsic motivation as a key ingredient for building world models autonomously.
https://arxiv.org/abs/2611.13113
Vision-language models (VLMs) exhibit strong zero-shot generalization but remain vulnerable to adversarial attacks. Existing adversarial fine-tuning methods often use pre-trained model guidance to preserve image--text alignment while improving robustness. Our analysis of visual representation drift reveals complementary limitations of two representative forms of guidance. Matching visual features constrains all directions equally, providing broad coverage without semantic emphasis. Matching prediction distributions prioritizes semantically relevant directions, but remains limited by the candidate texts and prediction probabilities. These limitations can hinder preservation of pre-trained image--text relationships and degrade zero-shot performance. We therefore propose Textual Response-Aware Guidance (TRAG), which constrains visual representation drift using the pre-trained text encoder's responses to local variations in input embeddings. TRAG constructs anchor-level constraints from the response directions and magnitudes around semantic anchors, then aggregates them using pre-trained image--anchor relevance, yielding an image-adaptive constraint. Extensive experiments across architectures and tasks demonstrate state-of-the-art clean and robust performance. Further analysis shows that TRAG helps preserve pre-trained image--text relationships.
https://arxiv.org/abs/2610.40309
Vision-language-action (VLA) policies have demonstrated strong generalization in robotic manipulation, yet processing dense visual tokens often incurs high inference latency. Visual token pruning offers a practical way to reduce this cost, but identifying a compact set of tokens that preserves reliable action generation remains challenging. Existing methods largely rely on static or immediate heuristic cues to estimate token importance. We introduce APS-VLA, a framework that learns visual token selection from robot demonstrations while keeping the pretrained VLA backbone frozen. A lightweight Decision Influence Estimator scores tokens conditioned on current visual observations, language instructions, and robot proprioception. Its training combines two complementary signals: action supervision matches both the full-input policy predictions and demonstrated actions, while an auxiliary visual transition objective uses demonstrated actions to predict subsequent visual features from the selected representation. At deployment, the estimator performs a single hard selection and packing step before the policy's native action-generation procedure; the auxiliary predictor is removed. Experiments with OpenVLA-OFT on LIBERO and with $\pi_{0.5}$ and XR-1 on RoboCasa365 demonstrate inference acceleration with modest changes in task success. Across five real-world tasks on two dual-arm platforms, APS-VLA reduces mean policy inference latency from 137.25 to 93.25 ms, achieving a $1.47\times$ speedup with an average success rate of 61%, compared with 60% for the unpruned baseline. Together, these results support APS-VLA as a modular, plug-in framework for efficient robotic control across VLA architectures, preserving the underlying policy parameters and native action-generation procedure.
https://arxiv.org/abs/2611.13030
Our closed-track entry (Arena username Vexorium-Agent) scored 86.31 points on the final board (L = 77.07, H = 81.86, Z = 100) over the 60 final problems, against 33.33 for the reference proofs, 8th at the deadline (top: 92.57), for US\$166.33 at list prices with every call counted (per problem: mean US\$2.772, at most US\$2.874; US\$3 cap). Three practices carry the entry. First, each paid stage runs the model that measured best per dollar for it among the 9 we tested in our own harness: GPT-6.1 Sol the early sessions, GPT-6.1 Terra the later ones and the tail at 70% of Sol's cost per turn, and Gemini 4 Argon the review and repair; Sol also runs the style rewrite. Second, what our development and research runs learned ships as compiler-checked programmatic rewrite rules rather than prompt text (+0.93 points from rules fixed before the release; +0.26 and +0.30, in sample, from passes built or switched on after the release). Third, every call is recorded and reviewed before each change (all 9 board scores came within -0.063 to +0.015 points of our meter's prediction). The harness, gate and many of these rules also serve our open-track entry. We report the full bill and the spread of three runs.
https://arxiv.org/abs/2611.13011