Remote sensing disaster assessment is moving from simple damage classification toward comprehensive report-level analysis. In this setting, responders need not only damage descriptions but also verifiable evidence to guide response and recovery. Recent vision-language models (VLMs) and tool-augmented agents offer strong multimodal reasoning and perception, but they struggle to provide the evidence required for high-stakes verification. To address these limitations, we propose DisasterAgent, an evidence-driven agent that uses a Perception-Augmentation-Generation workflow to generate verifiable disaster reports from bitemporal optical imagery. DisasterAgent first invokes a specialized vision expert toolkit in the perception phase to build a Spatiotemporal Perception Graph (STPG), serving as a structured memory to store and organize disaster-induced changes across space and time. Evidence-based Augmentation (EA) then extracts evidence from the STPG, attaching inline citations that form a closed evidence loop from each statement to the supporting visual evidence. Extensive experiments on the DisasterM3 and RSCC benchmarks demonstrate that DisasterAgent consistently outperforms general-domain VLMs, remote sensing-specialized models, and existing remote sensing agents. Notably, compared with the original model, it improves the average score for disaster report generation on DisasterM3 by 13.5\% and ST5-SCS on RSCC by 15.34\%, providing a more accurate and verifiable solution for disaster emergency response. The code will be available at https://github.com/ZhijianLi2003/DisasterAgent.
3D reflection removal is crucial in applications demanding high-fidelity visualization, such as museum exhibitions. In this work, we propose Reflection-Free 3D Gaussian Splatting (RF3DGS), a novel method for superior reflection removal. RF3DGS exclusively leverages a widely-adopted photometric prior for reflection removal: that the minimum pixel intensity across multi-view images corresponding to the same 3D point strongly correlates with reflection-free information. However, generalizing this principle to a 3D Gaussian Splatting within a multi-view context is non-trivial. To address this, RF3DGS first optimizes a vanilla 3DGS model using a geometric prior to estimate a coarse, reflection-free depth map for each view. Leveraging these depth maps and photometric prior, a 3D warping process aggregates reflection-free information from auxiliary views to construct a clean transmission pseudo-label for a given reference view. This pseudo-label provides a robust self-supervised signal for optimization. We further introduce an Adaptive Twin Rendering (ATR) loss to synergistically utilize this pseudo-label while preserving high-frequency details from the original, reflection-degraded images. By adeptly fusing these internal 3D self-priors with the ATR loss, RF3DGS demonstrates state-of-the-art reflection removal performance across both synthetic and real-world scenarios. Furthermore, we establish the first 3D reflection removal benchmark, based on the Unity rendering engine, for rigorous quantitative evaluation.
Monocular depth foundation models make it tempting to replace depth sensors on mobile robots with a single RGB camera. We show that, used off the shelf, this is unsafe: in a simulated indoor navigation benchmark, a standard Nav2 stack driven by Depth Anything V2 (metric) collides in 63\% of 120 episodes, because the model systematically over-estimates the distance to nearby obstacles. We propose a lightweight plug-in module that turns predicted depth into \emph{safe depth}. A 0.58\,M-parameter head on top of the \emph{frozen} depth model predicts, per pixel, a log-depth correction $\mu$ and an uncertainty $\sigma$, trained with a $\beta$-weighted Gaussian negative log-likelihood. The navigation stack consumes the one-sided bound $\hat d_{\text{safe}}=d\,e^{\mu-k\sigma}$, where $k$ is set by conformal calibration on a held-out world, and uncertainty also decides which pixels may be treated as floor. The correction halves the depth error (AbsRel $0.262\rightarrow0.134$) and reduces near-range over-estimation from 59.8\% to 12.0\%. In navigation, safe depth raises the success rate from 37\% to 70\% and lowers collisions from 63\% to 3\%, close to the 2\% of ground-truth depth. A constant safety margin calibrated to the same pixel coverage reaches only 30\% success because it freezes the robot; uncertainty-aware safety wins 52 of 56 discordant paired episodes against it. Gains hold on glass, texture-less walls and dark regions.
Learning feature interactions is essential for accurate prediction and interpretable modeling, yet a high interaction score may reflect fitting noise rather than a reproducible pattern. Ranking methods score feature pairs but leave the cutoff to the user; testing methods set the cutoff but rely on repeated model refits or a model of the joint feature distribution. We present LENS, a model-agnostic framework for selecting feature interactions through independent replication. LENS trains two models with the same learning algorithm on disjoint data and accesses them only through predictions. For each feature pair, it removes main effects from the joint-effect surface to isolate an interaction pattern, ranks pairs by its strength, and tests whether the second model reproduces it at a meaningful strength. A single multiple-testing correction determines which pairs to retain, using only two model fits regardless of the number of features. We prove that a replication ratio alone can retain arbitrarily weak patterns shared by both fits, whereas a strength floor restores selection consistency, and we characterize when normalization in Friedman's $H^2$ reverses interaction rankings. On 16 synthetic designs, LENS matches leading model-specific ranking methods (e.g., TreeSHAP) and recovers 91% of material interactions, against 45% for the knockoff-based DIAMOND, without selecting false pairs; with 1,225 candidate pairs and no interactions, it makes substantially fewer false discoveries. On large-scale French motor insurance data, it identifies interactions documented in the actuarial literature and recovers 90% of the deviance gain obtained using all pairs. On two credit datasets, LENS selects few pairs while maintaining or slightly improving predictive performance over the additive baseline.
One motivation for modular neural networks is to divide computation across components that learn distinct, reusable functions, which could then be recombined to support compositional generalisation. In practice, modular structure alone does not guarantee that such roles develop, and unrestricted communication between modules can undermine any specialisation that does emerge. This raises the question of how modular communication should be organised. The problem can be viewed as an analogue of the perceptual binding problem, where synchrony has been proposed as a mechanism for coordinating distributed representations. We introduce SyncNet, a modular architecture that separates module content from a dynamical phase state: learnt Kuramoto dynamics evolve the phases, while phase alignment determines communication through a shared channel. We evaluate SyncNet on Sort-of-CLEVR for relational reasoning and on SQOOP for systematic generalisation, where lower pair variety demands more recombination at test time. SyncNet gives the best held-out accuracy of the models compared at the three lowest pair varieties, by 5.1 percentage points at the hardest, and is competitive on Sort-of-CLEVR though below the strongest attention baselines. Test-time overrides show that performance depends on phases that evolve: fixed phases cost as much as no addressing at all, and permuting which module holds which region is nearly as damaging. A content-addressed variant trained from scratch nonetheless matches it, so the benefit appears to come from the modular decomposition rather than synchrony itself.
Weight-space programming allows language models to read a trained neural network's weights and edit them. This opens the door to ``training'' by reasoning rather than through gradient descent. Prior work has demonstrated its functionality and promise for game-playing networks and fixing specific failures in language models, even extending to diffusion models. In this work, we present its applicability to models trained on biosignal data, specifically EEG and ECG data. We show that frontier models are not only able to debug specific faults planted in biosignal-based models without being told what these faults are, but are even capable of writing all the weights of a network from scratch, given only the architecture, the task, and a minimal set of training subjects. In particular, in one set of experiments, a language model was able to match standard gradient-descent training using $6\times$ fewer training subjects, through reasoning alone and without any gradient descent step. Networks written by Claude Opus 5.5 also scored higher than classic hand-crafted EEG pipelines (alpha band power, and common spatial patterns with linear discriminant analysis) when built from the same 5 subjects. We also highlight some of the limitations, which include the inability of language models to improve already well-trained classification models, and their incapacity to reliably recognize, from the weights alone, the task a network was trained on. Overall, our results indicate that weight-space programming can be useful for small networks, such as those deployed on edge devices and wearables, but the concept still requires further investigation before it can be used in practice.
Quantum hardware can provide feedback for training circuit generators beyond the reach of exact classical simulation. We train generative quantum eigensolver (GQE) policies, transformers conditioned on molecular descriptors, to generate the state-preparation circuits for sample-based quantum diagonalization (SQD). Training runs first on a noiseless simulator and then with the IBM Heron processor ibm_kobe inside the loop, at 46 and 70 qubits ($1.2 \times 10^8$ shots for the four main training runs). A classical CCSD(T) reference energy of the training molecules enters only an admission gate and two auxiliary reward terms. During hardware training, a reward penalty prevents collapse to circuits that sample only the Hartree-Fock determinant. Policies trained on transition-metal complexes are applied without retraining to Sn(OH)$_3$-Me and its hydrates, minimal models of organotin photoresists for extreme-ultraviolet lithography, from 32 to 70 qubits. After one round of hardware sampling, followed by classical ranking and recovery of the cached counts with the molecular Hamiltonian and no further QPU calls, 27 of 28 runs at 46-70 qubits are within 15 mHa of the density-matrix renormalization group (DMRG) reference (median 4.5 mHa), and in all eight runs with an untrained control the trained policy is more accurate. At 56-70 qubits the selected circuits use 34-72 times fewer two-qubit gates than the local unitary cluster Jastrow (LUCJ) circuit transpiled for the same device. These results demonstrate that circuit generators can be trained with quantum hardware at 70 qubits.
OCR post-processing needs OCR text aligned word by word with its ground truth. Data sets rarely provide that mapping, and a monotone aligner can lose its place when zoning errors permute the reading order. We revisit this problem, the Zoning Problem of our 2019 paper, with a protocol that tests correctness under such permutations. We apply OCR-side reading-order permutations with construction-time word-level gold and score correct-match F1, cascade length and recovery. We compare the 2019 location-aware aligner, a re-anchoring extension of it, a RETAS-style baseline (RETAS-lite), a multi-region aligner that combines unique-word and unique n-gram anchors with chunk-to-window assignment, genalog's RETAS and the original Java RETAS. Experiments use synthetic character noise with block moves, paragraph swaps and column interleaving, historical newspaper reading-order permutations, geometry-derived gold for Tesseract OCR, and a small word-box set. Synthetic tests show where local tracking fails after large displacements and where unique-anchor methods collapse when unique anchors are scarce. On our synthetic boilerplate page the 2019 method had the highest F1 among those tested. Historical newspaper and geometry-derived evaluations carry the limitations of approximate gold and the article-order convention. The first multi-region version has a quadratic scaling cost. Local tracking remains fragile under large reading-order changes; the tested anchor-based methods handle many of those cases but fail when unique anchors are scarce. Approximate gold, the article-order convention, missing residual annotation, and the lack of modern multi-column word-level ground truth remain limitations.
Counterfactual editing of single cells, predicting how an observed cell would look had one factor been different, enables in-silico screening where experiments are infeasible, but is useful only if the edit is faithful to the individual and not merely to the condition. Concept bottleneck generative models make the effects of interventions interpretable by decomposing a cell's representation into known and unknown factors. Previous work uses a linear cross-covariance penalty to disentangle known and unknown dimensions. We demonstrate that this first-order criterion does not fully separate information when non-linear interactions are present. We first demonstrate the inadequacy of linear cross-factor penalization with simulation experiments with controlled edits and known non-linear covariate effects. We then demonstrate that adding a nonlinear kernel criterion, the normalized Hilbert-Schmidt Independence Criterion, is able to recover the true intervention effects on simulation data. Finally, we apply the non-linear concept bottleneck model to a complex single-cell RNA-seq dataset and find that it improves the ability to generate counterfactual expression profiles of cells under different tissue environments.
Emotion vectors are directions in a language model's internal representations that are associated with emotional concepts. Prior work shows that these vectors can influence model behavior, but their role in sentiment classification remains unclear. This study investigates emotion directions in Qwen2.5-0.5B-Instruct and Qwen2.5-3B-Instruct. Eight directions extracted from GoEmotions are evaluated through sentiment probing, activation interventions, and supervised correction. Emotion-only probes achieve 88.66% and 90.70% accuracy on binary TweetEval. In exploratory matched-ablation experiments, raw emotion-space removal causes 2.39 and 1.10 percentage points more accuracy loss than matched isotropic random-subspace removal for 0.5B and 3B, respectively. All four TweetEval raw and denoised effects are significant after Holm adjustment across eight comparisons. An exploratory SST-2 extension shows smaller raw excess losses of 0.66 and 0.57 points, with no effects surviving Holm adjustment. Frozen sentiment-correction classifiers are also tested on 20,000 Amazon Polarity reviews. Emotion-based correction improves accuracy over uncorrected Qwen, but does not consistently outperform margin-only baselines. These results distinguish sentiment information available in model representations, the effects of changing those representations, and their practical value for classification. They do not establish a uniquely emotional mechanism.
Plug-and-Play (PnP) methods with generative priors have achieved strong performance in inverse problems, with recent flow matching models enabling fast and scalable reconstruction. However, a key challenge in PnP methods is maintaining consistency between data fidelity updates and the learned generative prior, as data-consistency steps can move iterates away from the flow trajectory. We propose an ADMM-based PnP framework for flow matching priors that enforces stronger consistency between the measurement model and the prior through auxiliary variables and dual updates. In addition, we introduce a trajectory correction mechanism that restores consistency with the flow via a combination of deterministic and stochastic projections. Our method provides a more principled integration of flow priors within PnP frameworks, achieving high-quality reconstructions with reduced computational cost.
Laser scanning has been utilized for decades to measure tree structure and geometry. Despite continuous advances in sensor technology, uncertainties arising from scanner characteristics and measurement conditions has remained relevant sources of error. In particular, factors such as laser beam diameter can introduce both systematic and random errors in the reconstructed tree geometry. This study investigates the main sources of these errors in mobile laser scanning (MLS) and quantifies their effects on stem curve (stem diameter) estimation. We discovered through simulations and experiments that the heteroscedasticity (difference in variance in ranging and scan angle) in the measurement noise is a key factor along with the beam width of the lidar, which cause the bias of the diameter estimates to depend not only on distance, but also on the tree diameter. We demonstrated that this bias can be reliably predicted and negated for trees with a diameter at the breast height greater than 10 cm using technical specifications of the sensor without any additional calibration. In experiments, the proposed bias compensation method reduced the root-mean-square error of stem diameter estimates to sub-centimeter level on boreal forest test sites. However, for small trees with diameter below 10 cm, the diameter bias becomes sensitive to the diameter of the tree$\unicode{x2013}$an unknown factor$\unicode{x2013}$and other factors, such as different bark and out-of-roundness of the stem, limiting the precision of lidar for measuring geometry of trees and applicability of the presented method.
Optimized floating-point inference is not reproducible across environments: the same model on different hardware, compilers or thread counts can produce different bits. We present tritify, an open-source runtime that executes the forward pass of real models -- decoder-only language models, Mixture-of-Experts decoders, and the Segment Anything 2.1 image/video stack -- with the entire run-time path in exact integer arithmetic; floating point is confined to offline quantization and the float reference used for verification. We make three separate claims and verify each. (1) Cross-hardware reproducibility: the integer path produces bit-identical output run to run, across CPU instruction-set architectures, and across CPU/GPU, carried by a reproducing SHA-256 of the output. (2) Decision fidelity: the integer path reproduces the float reference's decision on the models and inputs tested, with the numeric exceptions measured; per-layer fidelity through sixteen layers of OLMoE-1B-7B and of a 106-billion-parameter MoE shows no fidelity cliff. (3) Location independence: a model's computation can be split across instruction sets, CPU and GPU, and networked machines with zero effect on the output bits. GLM-4.5-Air (106,852,251,264 parameters, 427 GB of integer codes) generated its answer across three machines, none of which can hold its integer codes in RAM, bit-identically to a single-host reference. The consequence is verifiable distributed inference: an untrusted server or a fleet of devices runs the model while an open scalar oracle re-derives the exact bytes; a guarantee against a dishonest server is future work. We do not claim a per-device speed-up, bit-equality of floating-valued intermediates, or confidentiality. Code: https://github.com/pelloran-inc/tritify
Existing channel knowledge map (CKM) construction assumes quasi-static environments, failing when channels rapidly decorrelate. This paper investigates continuous online dynamic CKM tracking over an airborne 2D plane via a UAV fleet. We formulate a long-term error minimization problem under UAV mobility and reporting constraints. This problem, however, is intractable with unobservable errors for unmeasured cells. To tackle this, we propose a metric-driven scheduling approach featuring the Weighted-Age of Information (W-AoI). W-AoI evaluates cells by combining exact temporal staleness with spatial uncertainty predicted by a 2D U-Net. We prove that top-$K$ W-AoI selection maximizes the one-slot error reduction bound. Guided by this guarantee, we develop a lightweight greedy trajectory policy incorporating continuous along-path sensing. Simulations demonstrate our method reduces steady-state error by 3.8 dB$^2$ and 4.8 dB$^2$ over AoI-only and FIFO baselines, closely approaching an idealized oracle.
Test-time reinforcement learning enables language models to adapt using rewards derived from their own answers. However, majority voting can reinforce incorrect solutions even when correct alternatives are available. Our training audit shows that false positives (incorrect responses receiving positive rewards) dominate reward errors. We focus on a practical regime where ground-truth rewards are difficult to obtain at scale, while a limited labeled source set is available. We introduce TRUER, a framework that uses limited supervision to train a reward rectifier that distinguishes correctness from consensus. To achieve this goal, we train the rectifier with an emphasis on incorrect majorities and correct minorities, targeting cases where consensus and correctness disagree. After that, we freeze the rectifier and use it to correct rewards during subsequent test-time adaptation. By improving reward accuracy and thereby model performance, TRUER consistently outperforms naive TTRL in both metrics across all evaluated configurations.
Recent advances in large language models (LLMs) and autonomous agents are creating new opportunities for hardware design optimization. This paper examines three connected directions toward autonomous RTL and synthesis optimization: (1) Realistic benchmarking. We discuss reference-based and open-ended RTL optimization benchmarks that evaluate optimization opportunities under realistic EDA workflows. (2) Tool-grounded RTL optimization. We present Dr. RTL, a multi-agent framework that automates critical-path analysis, RTL rewriting, and implementation-level evaluation, while continuously improving through reusable optimization skills. Evaluated on 20 real-world RTL designs, Dr.~RTL achieves average WNS/TNS improvements of 21\%/17\% with a 6\% area reduction over commercial synthesis. (3) Design-specific synthesis optimization. We explore agent-driven synthesis script tuning and the discovery of executable synthesis algorithms across intermediate representations, technology-independent optimization, and technology mapping. Finally, we envision AI-native design automation that jointly explores architectural, RTL, and EDA algorithm design spaces, guided by verification constraints and implementation feedback.
Deep Interactive Reinforcement Learning (Deep IntRL) accelerates reinforcement learning training by incorporating external advice. However, conventional Deep IntRL uses each piece of advice only in the current step and discards it after a single interaction. Persistent Advising and Broad-Persistent Advising (BPA) extend this process by storing advice in persistent memory and reusing it in similar states through state generalisation based on K-means clustering. This research proposes Encoder BPA, which employs a latent encoder to retrieve and reuse advice. In this framework, a trained encoder is used to map a state sequence into a latent representation at every time step. Advice can be retrieved through latent comparison, together with a top-(k) aggregation and a consistency condition to improve reuse stability. Experimental results show that Encoder BPA improves learning efficiency and advice utilisation at an early stage of training compared with different Deep IntRL approaches, especially in complex and partially observable environments.
Video overlay removal recovers the scene beneath burned-in subtitles, karaoke lyrics, titles, credits, logos and watermarks. Generative video models now fill a known region convincingly, but in practice the region comes from a detector or a segmenter, and it is wrong in both directions: it misses strokes, halos, whole frames and whole objects, and it covers background between characters, around boxes and on genuine scene text. Erasers that treat the mask as a hole, regenerating inside and copying outside, inherit every one of these errors. We present HintErase, which treats localization as a hint, not a hole. The model sees the full, unmasked video together with a hint channel; during training the hint is corrupted the way real localizers fail, while supervision follows the true overlay. The result is a single model that removes overlay the hint missed, keeps background the hint wrongly covered, and is trained to keep genuine scene text even when the hint lands on it. For full-length releases, a streaming pipeline handles interlaced sources, shot cuts, missed and jittering hints, paste-back and audio, with memory independent of the length of the video. This report describes the approach, the pipeline that turns it into full-length deliveries, and qualitative results; video results, with input, hint and output side by side, are on the project page.
Manifold embedding for data visualisation is dominated by a small set of hand-designed state-of-the-art (SOTA) algorithms. Their design targets are fixed at development time, resulting in inflexible algorithmic behaviour that cannot be readily steered towards user-specified structural priorities. Hand-designing new embedding algorithms to address user priority requires highly specialised expertise and is time consuming. To address autonomous algorithm design tailored to user preference, we introduce an agentic algorithm generation pipeline AutoManifold. It composes new manifold embedding algorithms from a constrained vocabulary of affinity, cost, and optimisation primitives, supported by multi-agent large-language-model (LLM) orchestration. AutoManifold conditions every stage of its design on user-specified structural preservation preferences, and iteratively refines the algorithm configuration through an LLM-guided, metric-grounded iterative loop. We compare the generated algorithms against the three strongest and most frequently used SOTA methods (t-SNE, UMAP, and PaCMAP) on real-world datasets spanning different difficulty regimes under identical evaluation infrastructure. The results demonstrate that LLM-driven, priority-conditioned algorithm synthesis can move beyond hyperparameter tuning, producing genuinely new algorithms that outperform traditional hand-designed methods in their ability to steer towards user-specified properties without compromising the intrinsic structure in the original data. The code of AutoManifold is available at https://github.com/buraksusam/AutoManifold_NeurIPS2026.
Conventional federated learning relies on the shared model backbone to aggregate client updates, limiting collaboration among institutions that deploy large language models (LLMs) from different families and at different scales. We propose Federated Verifier Pool (FEDVP), a new federated learning paradigm that enables heterogeneous clients to collaborate through shared verification rather than parameter aggregation. Each client trains a verifier using local data and responses generated by its chosen LLM backbone. The server pools these independently trained verifiers and combines their judgments to select answers without merging model parameters. This allows both LLM backbones and verifiers to differ in architecture and size, while keeping the LLM backbones frozen and deriving collaborative gains from improved answer selection. We evaluate FEDVP across five public medical question-answering datasets treated as federated clients, using Qwen and LLaMA backbones ranging from 3B to 14B parameters. FEDVP consistently improves performance across all five clients and outperforms the strongest conventional federated baselines by 1.46-5.79 percentage points in mean accuracy.