Retinal disease monitoring with optical coherence tomography angiography (OCTA) relies on segmentation of vessels and the foveal avascular zone (FAZ). Promptable models such as SEEM (Segment Everything Everywhere All at Once) allow click-based refinement, but without adaptation SEEM segments thin vessels poorly and needs many clicks. Fully fine-tuning its 342 million parameters on a few hundred images is costly and risks overfitting, and parameter-efficient methods such as low-rank adaptation (LoRA) usually adapt only the encoder. Yet SEEM reads every candidate mask from one feature map at quarter resolution, too coarse for vessels one to three pixels wide. This study introduces the Read-out Basis Adapter (RBA), whose read-out branch generates 32 image-specific basis maps at half resolution. Each candidate accesses them through its own vector and a zero-initialised projection, so the model starts exactly as SEEM. The encoder branch, a standard rank-1 LoRA, supplies the information needed to tell arteries from veins. RBA adds 0.47 million trainable parameters. Under simulated error-driven clicks, with the ground truth used only to place them, RBA outperforms LoRA, Conv-LoRA andAdaptFormer on OCTA-500 3mmlarge vessels. Averaged over three seeds, it reaches an intersection over union (IoU) of 85.16% and a Dice score of 91.93%, 2.90 and 1.72 percentage points above Conv-LoRA. On the 6 mm subset, one run gives margins of 4.35 and 2.76 points. The gain extends to arteries, veins and ROSE-O but not to the FAZ. In SEEM, adaptation capacity pays more at the mask read-out than in the encoder.
The Adam optimizer is widely used to train neural networks with highly nonlinear and nonconvex objective functions. Remarkably, Adam often drives the empirical training loss nearly to zero. This phenomenon has been partially explained by neural tangent kernel (NTK) analyses of gradient descent (GD) and stochastic gradient descent (SGD). However, analyzing Adam presents additional challenges. Its first-moment momentum couples stochastic gradients across iterations, while its adaptive second-moment preconditioner depends on the same stochastic gradient used in the update. In this paper, we establish linear convergence guarantees for Adam in training sufficiently wide two-layer neural networks with either full-batch or minibatch gradients. For full-batch Adam, we construct a Lyapunov function to control the effect of historical momentum. For minibatch Adam, our analysis builds on an exact virtual-iterate representation, a conditional analysis of the dependence between stochastic gradients and adaptive preconditioners, and uniform concentration bounds for the second-moment estimates. We prove geometric decay of the empirical loss for full-batch Adam and geometric decay in restricted expectation on a high-probability event for minibatch Adam. Beyond regression, we extend our convergence analysis to physics-informed neural networks (PINNs) for solving the Poisson equation, where differential operators and separate sampling of interior and boundary points introduce additional challenges. We establish analogous global convergence guarantees for both deterministic and stochastic Adam under explicit conditions on network width and Adam's parameters, characterizing the effects of minibatch size on convergence. These results help bridge the gap between Adam's empirical success and its theoretical understanding, providing sharper convergence guarantees for PINNs than existing analyses.
Machine unlearning in large language models (LLMs) aims to remove specific learned knowledge, capabilities, or behaviors while preserving model utility. Existing methods have mostly focused on the design of unlearning objectives, while largely overlooking the intrinsic matrix structure of transformer weights. In particular, standard optimizers operate on independent scalar coordinates, thereby ignoring this matrix structure. We propose spectral unlearning, an update framework that combines standard unlearning objectives with matrix-structured optimization. We show that the spectral update corresponds to the steepest-descent direction for transformers composed of linear layers with normalized inputs and outputs. Our approach is applied on top of existing unlearning methods and evaluated across complementary axes of memorization, privacy, and utility, summarized via their harmonic mean. Spectral unlearning consistently outperforms a broad range of baselines and closes most of the gap to a retain-trained reference. Code is available at https://github.com/unlearn-ml/spectral-unlearning/.
Large Reasoning Models (LRMs) produce explicit chains of thought to improve transparency, but they frequently overthink, generating long yet inefficient reasoning chains that inflate computational costs and risk introducing errors. While numerous reinforcement learning methods struggle to penalize verbosity to achieve conciseness, they typically apply the penalty uniformly across lengthy sequences, which leads to credit confounding, that is, the failure to distinguish essential reasoning steps from superfluous ones. In this paper, we study a measurable proxy for this confounding: high-entropy tokens that often coincide with surface transition points where reasoning re-enters verification. Based on this insight, we propose Adaptive Entropy-Sparing (AES), which selectively spares high-entropy tokens that occur before an efficient in-group reference, while penalizing overlength high-entropy suffix tokens associated with superfluous branching. Our AES thereby navigates the suppression-exploration trade-off, reducing overthinking while preserving average reasoning accuracy. We also present an entropy-length model that motivates why local uncertainty can serve as a proxy for future reasoning cost. Across eight reasoning benchmarks, AES reduces reasoning length by up to 52.9% while increasing mean accuracy by up to 3.8% over the base models, consistently outperforming existing efficiency-oriented methods. Our code is available at https://github.com/MobiusDai/AES.
Random tests of stateful code check laws, like reading back a written value, at states that random writes reach. The laws hold at every state exactly when the code behaves like a record with hidden data, so a pass is taken as evidence. This can fail: the laws can hold at every sample while the code is far from any record. The laws are local, but behaving like a record is global. Self-testing and stability theorems bridge this gap when the query distribution is controlled by the tester or preserved by the tested maps, but here it is generated by the tested writes, which need not be invertible. We ask how far a test must explore. An interface has $m$ binary variables, each with a read and two writes; in a record, reads return their bits and writes set only their bits. Samples are the ends of geometric runs of mean $\ell$ of uniformly random writes from a start distribution; a tester adds its own write chains of depth $d$. The distance is the least, over records with hidden data, of the largest probability that a read or write differs from it at a sample. From one start state, law checks at $O(m^2\log m/\epsilon)$ samples certify, at fixed confidence, any threshold $\epsilon$ after geometric runs of every positive mean, while runs of fixed positive length can pass at distance $1/2$. For the worst start distribution, take $\epsilon=m^{-\beta}$ with fixed $0<\beta\le1$. For $d\le cm$ with fixed $c<1/2$, a polynomial budget of samples and writes exists exactly when $\ell=\Omega(m)$, where law checks suffice. With no runs, the least depth for this budget is $(1/2+o(1))m$. Counting samples alone, in the worst case at $\epsilon=1/(32m)$ and $d\le m/16$, the critical mean for a polynomial budget is $\Theta(m\log(m/d)/\log m)$, by exhaustive search. So one start state needs only law checks, and the worst case at these thresholds needs exploration linear in $m$ if writes are counted.
Distribution Matching Distillation~(DMD) enables strong few-step diffusion generation by matching the student distribution to a pretrained teacher. Existing improvements to DMD often augment its supervision or modify its optimization machinery. We ask a complementary question: \emph{can DMD diagnose its own updates using signals already produced during training?} DMD inherently computes the discrepancy between a pretrained teacher and a fake model tracking the student distribution, but conventionally uses it primarily to determine the distribution matching direction. We find that the magnitude of this teacher--fake disagreement provides additional diagnostic information. Across extensive experiments, large disagreement in the high-noise regime consistently coincides with substantially higher-energy DMD updates. This reveals a dual role of disagreement: it determines not only \emph{where} the student should move, but also provides an intrinsic signal of \emph{how aggressively} the update should be applied. Based on this observation, we introduce Disagreement-Aware Distribution Matching Distillation (DA-DMD), which activates only in the high-noise regime and continuously modulates the update strength according to teacher--fake disagreement. DA-DMD reuses quantities already available during DMD training and requires no additional model, evaluator, or inference-time modification. We show that it exactly recovers vanilla DMD outside the selected region, reduces update energy only through selected samples, and exhibits redescending influence under extreme disagreement. Across multiple diffusion backbones, DA-DMD consistently improves few-step generation at the same inference cost. On SD3.5-L, DA-DMD improves HPSv2 from $28.23$ to $30.93$, while achieving a $64.4\%$ average win rate over DMD2.
Large language models are widely used in e-commerce user simulation, providing a foundation for understanding users and predicting human decisions. However, existing simulators still struggle to achieve both personalization and behavioral realism. Most methods predict the next action directly from interaction histories and the current context, without accounting for the evolving mental world that underlies user behavior. Theory of Mind suggests that behavior is driven by mental states, including beliefs, uncertainties, and intentions. Users may interpret the same information in distinct ways, leading to divergent behaviors, while an individual's mental state evolves through interaction and shapes subsequent actions. Ignoring this process prevents simulators from faithfully capturing individual differences and realistic behavior patterns, and undermines its effectiveness in downstream applications. To address this, we propose ShopMind, a Mental World Model for e-commerce user simulation that captures how consumers interpret shopping information and how their mental states evolve, better reflecting individual users and real-world behavior patterns. We further use behavior-mental state pairs synthesized by ShopMind to train a Mental State Reconstruction (MSR) model, which infers consumers' latent mental states from observable behavior. Experiments show that ShopMind improves personalization by 18.0% and behavioral realism by 19.6% over the strongest alternatives. Moreover, MSR improves performance on consumption value measurement and CTR prediction, showing the utility of ShopMind for understanding and predicting real user behavior.
Autonomous language-model agents are increasingly deployed on discovery-driven tasks where signal is sparse and most retrieved information is irrelevant. Current benchmarks fail to evaluate this regime due to dense signals, cooperative environments, and ground truth typically being only a few tool calls away. We introduce Legion, a benchmark framework coupling a stochastic enterprise-telemetry simulator (Legion-Sim) with a tiered library of 74 expertly-crafted adversarial scenarios (Legion-Bench) and an evaluation protocol that scores investigation efficiency alongside detection accuracy. Scenarios are injected into freshly simulated companies, allowing independent control over confounder density and event volume. We test five agent architectures on two simulated companies and find that while recall is high (0.82-0.88) precision lags behind (0.61--0.72). They fail to find all of the indicators of compromise of the attack or mistake genuine indicators with attack related ones, which is critical in the context of cybersecurity investigation. Failures are consistent across reasoning tiers and persist across confounder density and event volume. We release Legion, the scenario library, performance on common agent architectures, and pipeline for reproducibility and extension.
Generalist robot policies often fail at one step of a task they can otherwise perform. We ask whether such a failure can be repaired by editing the policy's weights directly, without gradients, training data, or a change to the instruction. Told to put the cube then the banana in the bowl, a zero-shot $\pi_{0.5}$ vision-language-action policy in RoboLab releases the cube before it reaches the bowl. Replaying 21 of these decisions under rewordings shows that it binds "in the bowl" to the wrong object. Closed-form instruction distillation copies the per-layer effect of an explicit rewording into the language model's MLP weights. On 128 held-out seeds, success rises from 27 to 106 episodes, with 81 fixed and 2 broken ($p=7\times10^{-22}$), while an update of the same size in random directions leaves success unchanged. The edit also lifts two other tasks without detectable harm on the remaining 13, and a variant that stores 0.16% of the parameter count keeps most of its effect. Placed in the action expert or action head instead, the same distillation helps less or not at all. Our released tool reproduced the fix end to end on new held-out seeds, from 8/64 to 36/64. Edits aimed at perception, a missing grasp, contact physics, or the gripper threshold did not help, and of seven blind Claude Opus editors, only the one given the misread instruction produced an edit that passed held-out scoring. These results suggest that weight edits work best when a known-good input already exists to distil. Because reruns with unchanged weights flip about a third of paired outcomes, evaluations should pair seeds, rerun the unedited policy, and confirm gains on held-out seeds.
Street view imagery (SVI) is widely used to audit urban environments remotely. Automated assessments usually assume that each image shows the target building clearly. In practice, an image may lack task-relevant evidence or show a neighboring house. If this goes unchecked, a model's score cannot show whether it judged the correct building. We present a reproducible SVI processing pipeline that makes image suitability and target identity explicit and checkable. We apply it to vacant house assessment in South Korea, using 10,311 images of 1,609 detached-house parcels. Task-suitability filtering adapts its criteria to image composition and administrative area. This keeps coverage even across areas. While keeping as many images as a single uniform filter, our filtering left 101 more parcels with a usable image. Target-house alignment identifies the parcel's house from camera metadata and parcel boundaries. It needs no building footprints or additional training. Alignment matched human labels for 85.6% of images, compared with 79.8% for the best simple rule. In vacancy classification, filtering and alignment raised AUROC from 0.684 to 0.719, but marking a different house in the same image still gave 0.709. Grad-CAM showed that the model often looked at areas outside the target house. Image suitability, target identification, and model evidence are therefore separate problems, and scores alone do not resolve them. Recording which images were used and which house was selected lets reviewers verify each assessment and can support human virtual audits.
Answer probes applied to chain-of-thought hidden states are often read as evidence that a language model has already decided its final response. This reading overlooks a competing signal, the candidate the model is about to mention. We introduce emission-controlled dual-target probing (ECDP), which fits the same low-rank reader family separately to the imminent mention and to the final answer at identical positions, with target-specific nulls and thresholds fixed before any fit. On correct DecisionDepth (DD) rollouts in which a distractor is mentioned before the final answer, Qwen3.5-4B decodes the next mention at .966 (null .234) but the final answer only at .221 (null .186), an emission gap of .696 [.648, .744]; Llama-3.1-8B gives .575 [.514, .632]. This pattern is not universal. In a pre-specified evaluation of Qwen3-8B with thinking enabled on held-out MMLU questions, using a full-prefix attention reader over every thinking position, both targets are decodable and the final answer is read far more strongly (mention-minus-answer lift -.398 [-.459, -.336] at layer 22 over 467 questions). Model, task, positions, and reader change together between these settings, and an attempted matched comparison on DD produced no valid statistics because the imminent-mention probe did not train reliably, so we do not identify what reverses the order. On DD, planted-signal controls show that the reader recovers an injected final-answer signal at one tenth of full dose. Answer-candidate decodability is therefore target-relative and protocol-dependent: a strong answer readout does not by itself establish when a model commits, and a probe that fails to train is not evidence that the information is absent. ECDP is a practical control for distinguishing imminent text emission from evidence of a settled decision.
Securing patent protection across countries requires drafting claims separately for each national patent office, in the relevant language and according to its legal and stylistic conventions. Patent claims can therefore differ across languages and jurisdictions because of translation, target-office drafting practices, and prosecution-stage changes. We formulate cross-lingual patent drafting as generating plausible candidate claim sets for a specified target language, jurisdiction, and prosecution stage from a source claim set in another language. The task models jurisdiction- and stage-specific transformations beyond content-preserving translation. We construct a large-scale Japanese-English dataset by linking Japan Patent Office (JPO) and United States Patent and Trademark Office (USPTO) records through 884,179 Patent Cooperation Treaty (PCT)-derived patent families. Across six released configurations, these families yield 4,067,168 directional claim-set instances; four core configurations form the main benchmark. Commercial machine translation and zero- and few-shot prompting capture part of the task, while supervised adaptation provides the strongest overall baseline, showing that task-specific patterns can be learned from linked cross-jurisdictional claim sets.
Self-supervised learning in Vision Transformer models, such as DINO, enables robust visual perception without requiring labeled data by leveraging large-scale image datasets. However, these models lack a focusing mechanism during training, resulting in perceptual representations that are often ambiguous or overly similar. In this work, we introduce two visual focusing loss functions designed to establish correspondence between the model's attention and specific regions within images. Specifically, we leverage the multi-head attention mechanism within the model to selectively steer the model's focus, leading to the emergence of spatial attention head specialization, characterized by diverse and spatially focused perceptual capabilities without requiring annotations. Through both qualitative and quantitative evaluations, we demonstrate that our method substantially increases the spatial selectivity and diversity of attention heads, improving both interpretability and performance on dense prediction tasks such as zero-shot image segmentation.
One-step stochastic world models avoid the mode averaging of deterministic feature predictors without the tens of network evaluations per step that diffusion and flow matching require. The simplest such objective is nearest-of-m matching (conditional implicit maximum likelihood estimation, IMLE): draw m candidate futures and pull only the one closest to the observed outcome. Recent work conjectures that a small m suffices for feature-space world models and that a strictly proper energy score should replace the hard minimum. We test both claims with one-step world models on frozen DINOv2/v3 patch tokens, using benchmarks whose futures can be enumerated. Under mode-like uncertainty, nearest-of-m with m=16 covers the enumerated futures better than 50-step flow matching (at about 1/30 of its latency), one-step flow generators and an energy-score generator at matched parameter counts, and its rollouts stay on the token manifold where flow rollouts drift. One quantity governs the choice of m: the nearest-of-m objective is the expected distortion of a random codebook of m draws from the generator. This view yields three predictions, recorded before the corresponding runs: (i) memorization of the realized outcome in contexts that never repeat, whose strength is governed by the codebook's miss distortion I(m) rather than its miss probability, and which an energy-score generator does not escape by raising m; (ii) a flattening of mode probabilities with m, matching a known optimum within 0.01; and (iii) under-dispersion on continuous high-dimensional outcomes, which bounds the method on real robot video. Per-step matching on self-rollouts cuts long-horizon error by a third on a stochastic Push-T variant; an ablation separates its manifold and coverage benefits. Planning success does not improve despite accurate risk estimates, a negative result we report with its diagnosis.
Traffic state estimation (TSE) reconstructs traffic density at unmonitored locations from a sparse set of sensors. Data-driven models work well when sensors are dense but degrade sharply as coverage thins. We show that imposing the Lighthill-Whitham-Richards conservation law as a physics-informed prior makes TSE more data-efficient. On four real corridors (NGSIM, I-80, PeMS04, PeMS08), physics-informed estimators generalise to held-out sensors where data-driven models, including classical and deep kriging, overfit the observed ones. On the two shock-rich corridors the effect is a clear crossover: as more sensors are held out, data-driven models collapse while the physics-informed estimator degrades gracefully, leading the no-physics network by about 26-38% at 85-90% held out and the strongest classical kriging by about 10-21% at 90%. A matched smoothness prior does not reproduce the gain, and an extended Kalman filter on the same LWR model is outperformed on the shock corridors, so the advantage comes from how the conservation law is imposed. Finally, one loss-conditional PINN removes the per-corridor loss-weight search, matching a 20-point search at about 5x lower cost.
We study the minimax regret of full-information online decision making in extensive-form games. For prediction with expert advice, the optimal regret is $\Theta(\sqrt{T\log K})$, where $K$ is the number of experts, or equivalently, the number of normal-form actions. For treeplex strategy spaces, existing efficient algorithms, including KOMWU and online mirror descent with the weight-one dilated entropy regularizer, achieve regret $\mathcal{O}(\sqrt{T\log|\mathcal{V}|})$, where $\mathcal{V}$ is the set of reduced normal-form strategies. We prove that this logarithmic dependence is minimax optimal in the worst-case full-information reward model. The main difficulty is that pure plans in a treeplex are not independent experts: their rewards are correlated through the recursive structure of decisions and observations. To capture this correlation, we introduce tropical Gaussian formulae, recursive Gaussian expressions built from maxima and normalized sums which preserve the variance. The index count of such a formula matches the number of reduced normal-form continuation plans in the corresponding subtree. Our main analytic result shows that every log-balanced tropical Gaussian formula with index count $N$ has Gaussian width $\Omega(\sqrt{\log N})$. While a direct approach struggles to recover this lower bound, our proof uses a spherical entropy profile to measure scale-dependent separation among the Gaussian directions, and then applies Tur\'an's theorem and Sudakov minoration to obtain the width lower bound. Applying this result to treeplexes gives an oblivious hard distribution with valid transition outcome and Rademacher terminal rewards, establishing a regret lower bound of $\Omega(\sqrt{T\log|\mathcal{V}|})$. This matches known upper bounds up to universal constants and shows that the DilEnt/KOMWU dependence on the reduced normal-form complexity is unavoidable.
Attention heads in video Transformers vary greatly in functional importance. Understanding their contributions requires examining not only which heads are important, but also how they participate in collective computation within the network. We study this collective computation from two perspectives: the downstream use of head outputs and the composition of interactions across layers. Ablation analyses of four video Transformers show that high-importance heads identified on a single task also support other video understanding tasks and help preserve the representational structure that distinguishes different videos. Representational responses to head ablation correlate weakly with head importance after the attention module, but more strongly after the MLP in the same block. Adapting this downstream processing through MLP-only fine-tuning further improves the use of lower-importance head outputs. Meanwhile, through composition across layers, sets of heads with shorter spatiotemporal attention distances support stronger perturbation responses at distant tokens than sets with longer distances. Their joint ablation causes even greater performance loss than ablating equally sized sets of high-importance heads, revealing collective contributions not fully captured by individual importance rankings. These findings highlight collective computation as a key perspective for understanding attention head function: functional contributions depend on both how head outputs are used downstream and how local interactions compose across layers.
In modular grounding a frozen detector proposes regions and a vision-language model (VLM) selects one, at lower cost than letting the VLM decode a box, but it is unclear where such systems fail. We hold perception fixed---every trained selector chooses among the same frozen proposals, which contain the referent for at least $99.8\%$ of RefCOCO/+/g queries---and vary the selection interface one factor at a time for two VLMs of one family (Qwen2.5-VL-7B, Qwen3-VL-8B). On RefCOCOg, what the selector reads matters more than how it is built: the VLM's image-conditioned query encodes the referent's location and beats a text-only query by $0.15$--$0.18$ with every head tested (by less for InternVL3-8B, whose vision encoder differs; one head tested). The remaining errors are rankings among a few top candidates; a cross-attention selector supervised with every valid box reaches the VLM's coordinate decoding on RefCOCOg, and with Qwen2.5-VL on RefCOCO/+ when also trained on their data. Beyond that distribution it does not: on Ref-L4's Objects365 images, mostly of categories outside training, every trained head falls to $0.19$--$0.28$ while Qwen decoding keeps about $0.8$ (InternVL3's drops too), neither a wider detector prompt nor a class-agnostic proposer closes the gap, and on public reasoning benchmarks Ref-Adv-s and RefBench-PRO full-budget heads trail decoding by $0.19$--$0.37$ although proposals cover most referents. A trained verifier reduces to an existence probe on the VLM's query state and does not beat the VLM's yes/no answer on public no-target benchmarks.
When competing campaigns target a population of large language model (LLM) agents, observed opinion changes may reflect both campaign influence and changes that would occur without intervention. We study how the timing of campaign expenditure and the population's own evolution shape outcomes in a network of persona-conditioned LLM agents that exchange and revise opinions in natural language. A proponent and a challenger are tasked with generating opposing messages under equal delivery budgets, while a separate stance probe measures opinions without influencing the simulation. Across eight claims, we compare single-campaign and competitive conditions against matched no-campaign runs, and evaluate all nine combinations of early, sustained, and late spending on a reference topic. Baseline stance drift varies substantially across claims and, on some, accounts for a sizeable fraction of apparent campaign effects. After accounting for this drift, the direction of observed campaign-associated changes aligns more closely with the claim's epistemic standing than with the campaigner's assigned role. Uninfluenced populations also consistently lose stance diversity, demonstrating that directional opinion change and convergence are distinct phenomena. In networks with two opinion-segregated communities, campaigns primarily reduce differences between communities rather than diversity within them. LLM campaign agents do not always argue for their assigned positions, particularly when these conflict with established evidence. Finally, campaign outcomes depend on the opponent's expenditure schedule: unlike a matched numerical bounded-confidence population receiving interventions at the same times and locations, the LLM population exhibits a payoff structure with no pure-strategy equilibrium.
Submodular maximization seeks a small subset best representing a large dataset. In its federated setting, the server is a single point of failure and some clients may be Byzantine. Removing the server raises a trilemma of decentraliza- tion, efficiency, and security. We propose DeFSM, a Byzantine-robust and communication-efficient architecture that replaces the server with a lightweight smart contract and a pool of independent, consensus-free agents. Its aggregation rule, SafeGM, builds on two notions from combinatorial geometry: Tukey depth and safe points (the points lying in the convex hull of the good points wherever the bad points are). SafeGM makes the robust decision on 2D Gaussian random projections, which the contract can compute efficiently, and lifts it back as a convex combination of actual updates. Since lifting loses information, it uses the projected geometric median only when its depth certifies it, and otherwise a Tukey median. The share of bad updates is then bounded only by a count: how many there are, not what they contain. We prove, and confirm empirically, that this bound controls the solution quality. On CIFAR10 and PathMNIST, with up to 30% Byzantine clients and agents, DeFSM never collapses and retains on average about 91% of the quality of the ideal reference, the geometric median at a trusted central server. The geometric median decentralized in its best setting would retain about 96%, but the contract cannot compute it, and so each agent would have to download six times more data, or each client twice as much.