Computer Science (arXiv)

Regret matching+ (RM+) is parameter free, scale invariant, and central to large game solving, but its only general individual-regret guarantee grows as $\sqrt{T}$. A recent ICLR result used this envelope to prove that RM+ reaches an $ε$-stationary point of a smooth objective over a product of simplices in $O(ε^{-4})$ iterations, or $O(ε^{-8})$ from the standard zero initialization. We give an exact one-step conservation law for RM+. It states that forward utility gain pays for both squared state motion and growth of the regret-state norm. Norm growth is at most $\sqrt{m-1}$ times forward gain for $m$ actions, and the coefficient is sharp. This yields four results for unmodified RM+. Its regret on any utility path is controlled by centered temporal variation. Its regret is uniformly bounded under alternating play in every finite exact potential game, resolving an open question and making squared activation gaps summable. Both certified lazy and ordinary cyclic play attain an $ε^{-2}$ exponent. On any smooth, possibly nonconcave simplex objective, RM+ finds an $ε$-KKT point in $O(ε^{-2})$ iterations. Most broadly, for a smooth objective over an arbitrary product of simplices, cyclic block RM+ attains the same $O(ε^{-2})$ exponent from arbitrary initialization, with an explicit trajectory-dependent constant. The proof controls the finite objective loss caused by low-state blocks and then self-bounds every block state and the total squared path length. Complete proofs cover zero states, sharpness, common-profile stationarity, and robust gain dominance. Oracle-normalized diagnostics compare RM+ with predictive and smooth extra-gradient variants on graphical potential games and dense nonconvex objectives.

This paper addresses the multi-robot task allocation (MRTA) problem, which is essential for delivery and logistics applications. Our approach first defines a new cost function that transforms the MRTA into a unified optimization problem capturing both task assignment and routing. A bi-layer ant colony optimization (ACO) algorithm is then introduced, integrating two interdependent decision layers within a single colony process to solve the problem. This hierarchical framework enables simultaneous optimization of task allocation and route planning across multiple robots. Comparative experiments with mixed-integer linear programming (MILP) and particle swarm optimization (PSO) demonstrate that the proposed bi-layer ACO achieves the shortest total travel distance and fastest completion time across all task sizes. Specifically, it reduces total travel distance by up to 17.7% and completion time by nearly 20% compared with baseline methods. These results confirm the efficiency, scalability, and reliability of the proposed bi-layer ACO for multi-robot delivery tasks.

Differentially private training adds isotropic Gaussian noise to clipped gradients, corrupting every singular direction equally. In vision models, where spatial correlation concentrates gradient energy into a low-rank subspace, most of this noise falls in directions that carry little signal. Spectral gradient orthogonalization via polar decomposition is introduced as a post-processing step that recovers directional signal from the noisy gradient's low-rank structure at zero additional privacy cost. A phase transition governs the utility of this approach: orthogonalization improves accuracy only when the per-direction spectral signal-to-noise ratio (SNR) suffices for singular vector recovery; in low-SNR regimes, the directional bias of the gradient is replaced by a nearly random orthogonal update, and the transformation is harmful. The recovery threshold is determined by the spectral gap of the gradient and is surpassed at large batch sizes. Empirically, the benefit scales with model capacity: spectral orthogonalization achieves a +20.9% improvement over DP-SGD on WRN-28-10 (B = 4096) and +14.9% on ResNet-18, while reducing inter-run variance by a factor of two to three. In the fine-tuning regime, spectral orthogonalization matches the stability of DP-Adam while maintaining a first-order memory footprint. Combining spectral with temporal denoising yields 50.3% on CIFAR-10 (epsilon = 4), the highest accuracy in any tested configuration. These gains are specific to moderate-to-high-SNR regimes such as large-batch training of higher-capacity models. Small-batch or low-SNR settings are better served by DP-SGD or temporal denoising.

Chart editing requires inferring and modifying visualization code from a reference chart image based on an editing instruction, challenging fine-grained visual reasoning, instruction following, and executable code synthesis capabilities of MLLMs. Large reasoning models (LRMs) with extended Chain-of-Thought (CoT) reasoning are suitable for tackling such complex multimodal tasks. However, our preliminary study reveals an ``inverted-U'' relationship between reasoning length and chart-editing performance: Excessive reasoning often leads to ``overthinking,'' where models drift toward hallucinated visual details or get stuck in redundant reasoning loops. To address the gap, we introduce REChart, a two-stage training framework that provides process-level supervision over intermediate reasoning steps, improving both editing fidelity and reasoning efficiency. First, we synthesize 200k high-quality reasoning trajectories for supervised fine-tuning from a large image-instruction-code pool, using a role-specialized agentic Reason-Score-Refine workflow that iteratively refine the chart code toward higher quality. Second, we optimize the model via reinforcement learning with two complementary rewards: a \emph{fidelity} reward evaluating code correctness, visual fidelity, and structural consistency, and an \emph{efficiency} reward that assigns each rollout a random thinking budget, truncates the reasoning process, and credits the final reasoning segment according to its contribution to the output. On the ChartEdit and ChartMIMIC benchmarks, our model achieves state-of-the-art chart-editing performance among open-source models of comparable scale, while mitigating overthinking and reducing average reasoning token usage by 79.0\% under a maximum thinking budget of 16,384 tokens compared with the base model.

Group Relative Policy Optimization (GRPO) has become a widely used approach for post-training Large Language Models (LLMs) for reasoning. In GRPO, the group gradients induced by different queries within the same mini-batch are directly averaged to form the policy update. However, these group gradients can point in conflicting directions. Our empirical analysis suggests that group-gradient conflicts tend to be associated with less effective policy updates, motivating the need for a reliable aggregated update direction under such conflicts. Standard GRPO aggregation treats the realized group gradients as deterministic contributions and does not account for differences in their reliability during aggregation. To address this issue, we propose Gradient Uncertainty-Aware Policy Optimization (GUPO), which models each group gradient as a random variable under a Bayesian formulation and estimates its probability distribution. GUPO then derives gradient uncertainty using a Dirichlet-based formulation and uses it to calibrate the contribution of each group gradient during aggregation. Extensive experiments on multiple benchmarks demonstrate the effectiveness of GUPO.

Garages have a long tradition of tinkering, creativity and innovative change. School of Garage, a participatory artistic summer school project in Chemnitz, the European Capital of Culture 2025, took up this tradition and turned old Eastern Bloc garages into temporary ateliers for collaborative making and discussion. In our HackLab garage we conceptualized and created the Oracle of Chemnitz within one week. It gives a place filled with history back its stories. It is an interactive installation of artifacts from the past typically found in garages: an old typewriter, radio, desk, tires, mixer and a rotary-dial telephone. Each got a name, personality and story to tell. The phone rings when a visitor approaches. Once answered, it asks for name and month of birth before a story about a device is told, along with hints to other places in the city. Around 2,700 visitors interacted with the system over three months.

Vision encoders are a critical component of vision-language models, and scaling their capacity effectively improves performance. However, dense scaling increases compute cost and inference latency. Mixture-of-Experts (MoE) architectures offer a compelling alternative, having enabled efficient scaling in LLMs, yet the MoE design space for CLIP-style vision encoders remains underexplored at State-of-the-Art (SOTA) levels. In this work, we systematically study MoE designs for vision encoder scaling and find that fine-grained MoE topologies yield substantial gains over both dense and standard MoE counterparts. We further propose an auxiliary-loss-free balancing variant for better expert utilization, and design a specialized MoE kernel to mitigate inference latency overhead. To enhance video capabilities while preserving image knowledge, we introduce frame-level distillation paired with a novel freezing mechanism. We pretrain a series of Mixture-of-Experts Vision Encoders (MoE-ViE) across a range of sizes, all consistently outperforming their dense counterparts. Our largest model matches the zero-shot performance of a SOTA encoder 1.7x its size at 76% of its latency. When aligned with an LLM, MoE-ViE surpasses all compared encoders on image and video benchmarks, including those with up to 5x more activated parameters. Code is available at https://github.com/facebookresearch/moe_vie.

Modern software systems evolve through frequent commits that implement bug fixes, features, and security patches. Although code review and testing are widely used to check these changes, they often provide limited assurance for memory-safety issues. Code reviewers may miss subtle boundary, lifetime, or initialization errors, while existing tests may not exercise the specific paths affected by a commit. Fuzzing is effective at exposing such bugs, but applying it to every commit remains impractical because whole-program fuzzing is expensive, requires suitable harnesses, and may still fail to reach the code changed by a commit. In this paper, we introduce COMMITGUARD, a commit-aware differential slice-based fuzzing approach for verifying code changes. The key insight behind COMMITGUARD is that the pre-commit version of a modified function can serve as a behavioral baseline for interpreting bugs found after the commit. For each target commit, COMMITGUARD identifies modified functions, extracts compilable code slices from both the pre-commit and post-commit versions, and fuzzes the paired slices independently. It then compares sanitizer reports across the two versions and reports bugs that emerge only in the post-commit version as candidate commit-induced bugs. We evaluate COMMITGUARD on 300 commits from openSSL, libpcap and leptonica. Slice fuzzing initially produces 518 sanitizer reports across these commits. By comparing pre-commit and post-commit slices, COMMITGUARD narrows this large output to 7 candidate commit-induced bug reports that require manual triage. Manual validation confirms 5 of these reports as real bugs that were fixed by developers of the examined projects after we reported them, while only 2 reports were classified as false positives. COMMITGUARD analyzes a commit in 32.4 minutes on average and achieves 75.36% average coverage of modified functions.

Text which has been translated from another language tends to carry with it evidence of translation$\unicode{x2014}$hence, it is often referred to as $\textit{translationese}$. Multilingual large language models (MLLMs) generate text in a variety of languages. However, it is still unclear if MLLMs' generations resemble internal translation (from English or, potentially, other languages) and, thus, result in translationese. Here, we ask the following research questions: (1) Does text generated by MLLMs resemble translationese? (2) How does translationese produced by MLLMs differ from translationese produced through direct translation? We leverage established indicators of translated text to evaluate text generated by state-of-the-art MLLMs in five languages, comparing to both non-translated and human-written baselines in order to isolate translationese from other kinds of interference. Through the use of high-accuracy classification models, analyses of variance on individual linguistic features, and the collection of human annotations in a subset of two languages (German and Spanish), we assess the translationese content of MLLM generations and examine the key features that distinguish MLLM-generated text from typical translation-related interference.

Persistent clouds blind optical satellites during floods. While Synthetic Aperture Radar (SAR) penetrates clouds, its raw data is noisy and lacks clear contrast. To mitigate this, recent studies utilize deep learning models to translate SAR into cloud-free synthetic optical imagery for downstream tasks like water body segmentation. However, because raw SAR is the original source for both of these operations, a critical methodological dilemma arises: during complete overcast should segmentation models process the raw SAR directly, or rely on a translated synthetic Normalized Difference Water Index (NDWI) proxy? This study resolves the debate by demonstrating that synthetic NDWI yields better results, as the translation process acts as a powerful filter against radar noise. This raises a natural second question: what if we utilize both? Building on our findings, we introduce a Combined Framework that integrates both raw SAR and synthetic NDWI into a unified model. By fusing the sharp physical boundaries of raw SAR with the high contrast of synthetic NDWI, this hybrid approach consistently outperforms all standalone methods.

Modern software systems are increasingly complex, and static analysis tools are commonly used to identify potentially vulnerable code by issuing warnings. However, these warnings often require manual inspection to confirm whether the reported issues are real, making the process time-consuming and error-prone. Directed fuzzing has emerged as a powerful automated technique to validate the warnings. However, applying it to the entire project in response to each warning is computationally infeasible, often requiring days of execution to achieve only incremental improvements in code coverage. We present SNIPTEST, an execution-based warning triage framework that generates and fuzzes compiled code slices centered around static-analysis warnings. Rather than proving exploitability in the full program, SNIPTEST provides evidence about how a warning behaves under progressively expanded sliced execution contexts. It employs a layer-by-layer slicing strategy, incrementally expanding context around the target location to validate potential vulnerabilities with increasing precision. We evaluate SNIPTEST on a benchmark of 97 true vulnerabilities and 97 false alarms across three real-world projects. SNIPTEST produces Possible True Positive evidence for 53 of 97 confirmed vulnerabilities (54.6%) by triggering the corresponding bug oracle consistently across all three analyzed slice levels, while the remaining cases are unreachable. Particularly, in 40.2% of these cases, it exploits the vulnerability along the observed execution path, matching the top three stack frames. On the 97 confirmed false alarms, SNIPTEST produces Possible False Positive evidence for 54 cases (55.6%) by reaching the warning without triggering the bug oracle, but misclassifies 28 cases (28.8%),and the remaining cases are unreached. Finally, we demonstrate the practical relevance of SNIPTEST by identifying CVE-2025-11964.

Spiking neural networks (SNNs), characterized by bio-inspired neuronal dynamics and event-driven communication, have attained significant progress in recent years. Nevertheless, training deep SNNs remains challenging due to spatiotemporal information loss and gradient mismatching. To simultaneously address these issues, we propose a noisy group neuron (NGN) model, which incorporates population-level synchronous resetting and neural stochasticity as fundamental computational mechanisms. We then develop the NGN method as a framework that combines the NGN model with backpropagation learning based on mean-field dynamics. We demonstrate the advantages of the NGN method through theoretical analysis and experimental validation on CIFAR-10, CIFAR-100, Tiny-ImageNet, DVS-Gesture, N-Caltech101, and CIFAR10-DVS. The proposed approach achieves an accuracy of 87.35% on CIFAR10-DVS within 10 inference time steps. These results support NGN as a practical approach to high-performance neuromorphic computing.

Reinforcement learning for coding agents increasingly relies on long-running agent harnesses to manage tool integration, repository contexts, and execution feedback. However, the native execution environments of these harnesses are inherently misaligned with policy-gradient training: environmental crashes and reward hacking corrupt outcome signals, while train-inference discrepancies decouple rollout behavior from policy updates. To address this, we present LEGO-RL, a framework that bridges native coding-agent harnesses with scalable policy-gradient optimization without modifying their internal control flow. LEGO-RL is built upon three pillars: (1) faithful optimization via in-process LLM proxying that captures raw generation streams for token-level alignment and robust trainer-side log-probability recomputation, even under harness-side compaction or re-serialization; (2) reliable execution via scalable sandbox orchestration featuring image caching and stage-wise defenses to mitigate reward hacking; and (3) observable training through an integrated plugin that automates validation and monitoring, paired with a Live UI for granular trajectory diagnostics. We evaluate LEGO-RL by training the sparse MoE model Qwen3.5-35B-A3B with GSPO across three native coding-agent harnesses. LEGO-RL improves Qwen3.5-35B-A3B across OpenHands SDK (64.0% to 70.4%), Claude Code (62.4% to 68.2%), and OpenCode (57.2% to 66.6%) on SWE-bench Verified, while maintaining a rollout-training probability correlation above 0.99.

Terrain level-of-detail is measured almost exclusively on digital elevation models: single-valued, smooth at the sampling scale, sampled from real topography. Game terrain is often none of these. We compare six rendering methods - height-field ray marching, voxel-accelerated ray marching, sliced proxy geometry, per-sample bar rasterization, compute scattering, and a fitted triangle mesh - implemented in a single engine over a single data path, on the hand-authored multi-layer terrain of Vangers (1998), scored against a CPU ray cast of the same source data. Every method must preserve the two solid intervals available at a ground sample, render at interactive rates, and reflect local terrain destruction without reloading the level. These constraints rule out treating caves as decoration or amortizing a static preprocessing step over an immutable map. From the original game's top-down camera the six methods look interchangeable. At eye-level horizons they do not: point scattering loses coverage, slicing bands, and an over-simplified mesh can miss a wall. At the selected quality settings a greedy triangulated irregular network (TIN) has the lowest mean frame time on every device we measured, but the fit cost is set by the second layer rather than by floor relief, and making that mesh editable retains 319 MiB of GPU geometry and 535 MiB of CPU triangulation. All six implementations use the same native wgpu / WebGPU API and canonical WGSL. We release the engine, the harness, and a one-command measurement protocol.

Long-sequence 3D reconstruction from RGB videos requires both accurate local geometry and globally consistent camera motion. Feed-forward models provide strong depth and pose predictions, but their memory cost prevents joint inference over long sequences. Chunk-wise processing improves scalability, yet independently predicted chunks often exhibit scale drift, pose errors, and point-cloud misalignment. We present GeoWeaver, a unified framework comprising a Geometric Prior Model (GPM) and Test-Time Adaptation (TTA). The GPM predicts chunk-wise depth, confidence, and camera parameters as adjustable geometric priors. TTA then performs sequential initialization, global chunk-level Sim(3) alignment, and coarse-to-fine refinement of camera poses, affine depth corrections, and intrinsics. Dense correspondences provide adjacent, cross-chunk, and long-range constraints, while a robust CDF-style objective jointly optimizes weighted 2D reprojection and 3D consistency residuals. This design preserves local geometric accuracy while correcting accumulated pose, scale, depth, and calibration errors. Experiments across diverse long-sequence benchmarks demonstrate improved camera accuracy, global consistency, and point-cloud quality. Ablations verify the contribution of each adaptation stage, and applying the same TTA procedure to different geometric prior models consistently improves their trajectory estimates, demonstrating that GeoWeaver is not tied to a specific GPM.

Linearizability checkers check if an operation history, observed by concurrent clients, is linearizable. They are used in testing distributed storage systems, and use the classic Wing-Gong (WG) linearizability checking algorithm. In this paper, we generalize the WG algorithm to make linearizability checkers more versatile: we can check other non-transactional consistency guarantees, like ordered sequential consistency provided by Zookeeper. Equipped with this generalization, we can also check for system-specific consistency guarantees that introduce additional ordering constraints over operations in a history, as per the system's specification. Our experiments with 8 distributed storage systems show that checking for system-specific consistency guarantees is easy to realize, reduces false negatives in testing, helps debug consistency violations, can be up to 370x faster, and can scale to more concurrent clients within the same checking time budget. We report 6 new consistency violation bugs, out of which 5 could not be found with existing consistency checkers.

Foundation-model policies for robotic manipulation are advancing rapidly on task success, but rigorous evaluation of whether they succeed safely is still lacking. We introduce ManiGuard, a specification-grounded framework for evaluating and improving the safety of foundation-model manipulation, comprising the ManiGuard-Bench task suite and a paired safety-annotated trajectory-generation pipeline. ManiGuard-Bench organizes six contact-rich household task families into 200 locked base tasks along a skill $\times$ constraint taxonomy, with safety specified independently of task success. Each task is evaluated under one in-distribution and four single-axis out-of-distribution perturbations that hold the safety specification fixed, giving 1,000 locked scenarios. Every rollout is runtime-checked by LTL$_f$-grounded automaton monitors over physics-grounded predicates rather than learned classifiers or LLM judges, in simulation and on a physical Franka platform. The pipeline pairs an automated motion-planning generator with human teleoperation, annotated by the same per-step monitor, and directly supports safety-aware fine-tuning; we release 8,000 safety-annotated demonstrations, 40 per base task. Benchmarking zero-shot and fine-tuned VLAs across more than 23,000 rollouts, we find: (i) safety must be evaluated independently of task success, as 6-21% of successful rollouts violate the specification; (ii) fine-tuning on our suite raises safe task completion from near zero to 7.5-29.8% and engaged-and-safe behavior from 16-40% to 51-72%; but (iii) a gap remains that scaling demonstrations does not close, with 21-42% of engaged rollouts still violating, two of six families below 2% safe success for every policy, and these failures persisting under distribution shift and on hardware.

We show that the balancing weights technique of Li (2026) actually produces an approximate *pseudo-circulation* of a directed, capacitated graph in $m^{1+o(1)}$ time. Together with standard flow techniques, we obtain an $m^{1+o(1)}$ time maximum flow algorithm that avoids the interior-point method framework of recent almost-linear time algorithms (Chen et al. FOCS 2022, van den Brand et al. FOCS 2024).

Channel knowledge map (CKM) is a promising technology for environment-aware sixth-generation (6G) wireless networks. However, most existing CKMs are tightly coupled with wireless devices and downstream tasks, which limit their scalability and reusability in wireless networks. To address these limitations, this article proposes the concept of universal CKM (uCKM) as a foundational wireless environment prior, which aims to enable cross-device and cross-task channel knowledge transfer for environment-aware wireless networks. We first revisit the representative CKMs and discuss their limitations. Then, the uCKM-enabled new paradigm for environment-aware wireless networks is introduced, and its benefits are highlighted from the perspectives of uCKM construction and utilization phases, for which we propose the visions of ``All for uCKM'' and ``uCKM for All'', i.e., the data acquired by all devices and tasks should contribute to the construction of uCKM, and vice versa. Subsequently, we discuss the main challenges of uCKM and propose potential solutions. Last, we provide simulation results to demonstrate the feasibility and performance gains brought by uCKM and outline future research directions.

Biomolecular design underpins applications from molecular recognition to therapeutics and synthetic biology, yet de novo interaction design remains challenging-especially for DNA/RNA, underexplored non-protein modalities with scarce, heterogeneous complex data and sharper geometric and chemical constraints. We introduce MCTH (Monte Carlo Tree Hallucination), an inference-only framework that casts all-atom sequence-structure co-design as uncertainty-aware planning over hallucinated states from pretrained folding and inverse-folding models, with optional biophysical control within the same decision loop. MCTH treats these models as frozen black-box operators and uses Monte Carlo Tree Search to allocate a fixed inference budget across competing design trajectories, incorporating model confidence and uncertainty, as well as cross-expert consensus/disagreement when multiple predictors are available. Across protein-RNA, protein-DNA, protein-protein, and protein-ligand design, matched-budget experiments show that adaptive search improves over simpler sampling and cycling strategies, while held-out AlphaFold3 and Chai-1 evaluations demonstrate transfer beyond the search-time oracle. MCTH provides a shared planning layer across modalities while allowing task-specific folding, inverse-folding, and biophysical modules, requiring no fine-tuning or backpropagation through component models.

We introduce PTXBench, a benchmark for evaluating and adapting large language models (LLMs) to use architecture-specific PTX for GPU kernel optimization. PTXBench measures functional correctness, whether selected target instructions execute at runtime, and speedup over frontier libraries across GEMM and attention workloads on H100 and B200 GPUs. Our evaluation shows that architecture-specific PTX capability remains uneven: success rates fall substantially on complex attention backward workloads, and executing the target instructions does not necessarily translate into competitive performance. No evaluated model consistently matches frontier libraries across the suite. We further adapt Qwen3.6-27B using supervised fine-tuning. Repair-conditioned training improves several tasks, but generalization remains uneven; data coverage, balance, and the quality of the reasoning teacher matter in addition to dataset size. PTXBench provides an auditable testbed for measuring and improving LLMs' ability to exploit evolving GPU architectures.

Motivated by a conjecture of Vaikuntanathan and Zamir, we study the pseudo-mixing of Kac's walk on $\mathrm{SO}(n)$: whether short trajectories are indistinguishable from Haar measure by low-complexity tests. We prove that the first $k$ columns mix in Wasserstein distance in $O(n(k+\log n)\log n)$ steps for fixed accuracy, resolving a conjecture of Oliveira. Combining this with a representation-theoretic variance bound, we show that if $T=ω(nk(k+\log n)\log n)$, then every degree-$k$ polynomial normalized to have unit Haar variance has expectation under the $T$-step law within $o(1)$ of its Haar expectation. As an application, we show that this pseudo-mixing estimate can be used to prove the effectiveness of a fast Johnson--Lindenstrauss transform with the usual target dimension.

Sample efficiency is a central challenge in reinforcement learning (RL), particularly in image-based domains where agents must learn from high-dimensional visual inputs. Traditional sampling often relies on random or suboptimal experience selection, leading to redundant updates and slow learning. Improving efficiency requires mechanisms that prioritize informative experiences while also encouraging effective exploration. Prioritized Experience Replay (PER) addresses part of this challenge by reusing high-value transitions, while intrinsic rewards promote the exploration of novel or uncertain states. However, their integration has not been extensively studied. This paper introduces Novelty and Surprise Prioritized Experience Replay (NSPER), which uses novelty to capture underrepresented states and surprise to expose gaps in the agent's understanding of the environment. We further extend this with NSPER+R, integrating these signals as intrinsic rewards to jointly improve replay quality and exploration. Experiments on DeepMind Control Suite tasks show that NSPER and NSPER+R improve training efficiency and convergence speed compared to existing methods in image-based RL.

Generative models promise a route to explainable clinical AI: rather than probe a classifier, model the distributions of healthy and diseased patients and read explanations off the geometry between them. We build such a system - an optimal-transport rectified flow trained between two clinical distributions - and use it to ask a pointed question the field too rarely tests: do the resulting explanation heatmaps actually localize disease? On tabular tumour biomarkers (Breast Cancer Wisconsin) a single flow yields per-patient counterfactuals, an unsupervised malignancy score (AUROC 0.91; 0.93 +/- 0.01 across five seeds), and a label-free attribution that agrees with a supervised classifier (r ~ 0.5) - a compact, honest interpretability engine, though it never out-predicts logistic regression. Moving to chest X-rays, we show the transport heatmap is a population-level signal, not a localiser; a reconstruction-based, identity-preserving variant does localize synthetic lesions (pointing game 0.52), yet on real RSNA radiologist boxes it collapses to chance while only supervised Grad-CAM stays above it. The central result is a synthetic-to-real gap: label-free heatmaps that look compelling on planted lesions are not evidence of real localisation. We contribute a reusable optimal-transport recipe for generative explanations and a controlled benchmark for stress-testing whether they localize.

Merging finetuned models combines specialized capabilities without joint training or access to the original data. Most methods operate by linear arithmetic in Euclidean weight space, which cannot carry the geometry of the update. Orthogonal Model Merging (OrthoMerge) uses a single orthogonal transform for each weight matrix, but such a transform cannot change singular values. We propose CORAM, which partitions each target matrix into row slices, represents every expert slice by its singular value decomposition in the corresponding base-model SVD frame, and merges the task-specific factors on their corresponding manifolds. Because manifold averaging contracts the merged update, CORAM applies an amplification coefficient $λ=κ\hat{c}$. The scale c_hat is estimated from the expert and merged update norms and is approximately $\sqrt{N}$ for $N$ experts with comparable update magnitudes. The restoration strength kappa is selected from the dispersion of expert updates without evaluating candidate merged models. This rule remains within 0.72 points of the best swept value on all evaluated suites. CORAM also includes spread slicing to distribute highly updated rows across slices and a residual pathway for non-target layers. Across four suites covering three model families, 3B to 9B scales, and language and vision-language experts, CORAM improves over OrthoMerge by 0.25 to 1.35 points and matches or exceeds the strongest weight-space baselines.