Papers

LLM agents rely on retrieval tools to access external knowledge, yet visual agentic search remains severely bottlenecked by standard single-step retrievers. In current pipelines, the agent must issue text queries for every intermediate step, struggling when visual clues are difficult to describe or when the retriever fails to surface necessary intermediate evidence within its top results. We hypothesize that offloading multi-step navigation across the entire embedding space directly to the retrieval tool resolves this performance bottleneck. To study this systematically, we introduce VHOP, a flexible data generation framework and benchmark with five core difficulty levels testing both visual matching and search planning. Using this framework, we develop VHOP-Router, an end-to-end training pipeline—combining supervised fine-tuning, online imitation learning, and reinforcement learning—that transforms a standard embedding model into an autoregressive multi-step retriever. Operating directly in the visual latent space, VHOP-Router retrieves linked image chains in a single tool call without requiring the agent to formulate intermediate text queries. Experiments show VHOP-Router boosts retrieval performance from under 5% to 76.3%. In agentic search, it improves task success rates by 52.7% and reduces the average token length by 61% from 1886 to 728, whereas upgrading the agent yields only a 3.7% gain. Compared to a strong baseline where the agent retrieves the top 50 results per step, VHOP-Router maintains superior performance while reducing in-context images by $23\times$ and cutting the cumulative API payload by $35\times$. The models also generalize robustly to unseen difficulty levels and realistic test sets. Ultimately, VHOP and VHOP-Router provide an efficient and effective solution for visual agentic search that leaves native LLM capabilities entirely intact.

We propose ABrA-GD (Adaptive Bregman Accelerated Gradient Descent), an adaptive algorithm for relatively smooth convex optimization. The algorithm is derived from a computable primal–dual certificate that guarantees progress by comparing the objective value with a lower bound on the minimum of a regularized objective. This certificate enables adaptation to both smoothness and geometry, using only the relative strong convexity constant as a problem-dependent input. We introduce the dual Bregman Length Distortion Factor (BLDF), which measures how dual Bregman lengths change under an anchor shift or rescaling. Under bounded local dual BLDF, ABrA-GD achieves an accelerated $O(1/k^2)$ rate for convex objectives and an accelerated linear rate for relatively strongly convex objectives.

Desk-based learning and creative activities benefit from handwritten engagement. However, current generative AI tools deliver guidance through a separate screen, creating a gap between where users think and where assistance appears. To address this, in this work we design AIfred, a desk-based robotic arm with a projector mounted at the end-effector that places AI-generated guidance alongside handwritten work. AIfred combines workspace perception, context-aware content generation, and robot-mediated projection to support math assignments, image generation, and drawing tasks. In a user study (n = 36), we compared AIfred against ChatGPT (GPT-5.6 Luna) running on a laptop. Both tools performed comparably while assistance was available during the math assignment (6.7 vs. 7.3/10, p = .41), but AIfred improved short-term learning transfer by 60% once assistance was withdrawn (7.0 vs. 4.4/10, p = .003). In addition, independent art and design professors ranked drawings produced with AIfred better in 33 of 36 cases. Our findings indicate that spatially co-located AI assistance benefits tasks whose guidance shares a spatial frame with the work.

Searching in a linked list is one of the most basic problems in classical algorithms. Although the nodes of the list come with memory addresses, classically those addresses play no role in the cost of search: one simply starts at the head and follows successor pointers to search for an element. In this paper, we show that the quantum setting is different. Here, the ambient address space from which the list vertices are drawn can itself affect the query complexity. We study the following problem analogous to search in a linked list in the query complexity model: the input consists of an address universe $[N]$, a public start symbol $s$, a successor oracle $f$ whose non-$\perp$ values trace a hidden simple path $s \to a_1 \to a_2 \to \cdots \to a_\ell \to \perp$, and a marking oracle $g$ that marks at most one list vertex. The task is to decide whether the list contains a marked vertex. We prove that both the decision and search versions of this problem for all $N \ge \ell \ge 1$ have quantum query complexity $\Theta\!\bigl(\min\{\ell,(N\ell)^{1/4}\}\bigr)$. Thus, quite surprisingly, when $N < \ell^3$ the optimal quantum complexity is $(N\ell)^{1/4}$, which is strictly smaller than the $\Theta(\ell)$ cost of ordinary linked-list traversal. This gives a precise characterization of when the ambient address space yields a genuine quantum advantage for linked-list search. We extend our results and give the same tight asymptotic bounds for the natural double linked-list version as well.

Unused encoding states offer opportunities to simplify quantum circuits. For algorithms restricted to reachable subspaces, unspecified diagonal-operator entries can be optimized without changing ideal computation. We investigate exact diagonal completion for the quantum approximate optimization algorithm (QAOA) applied to placement, using permutation-preserving register swaps. We construct exact Manhattan-distance operators through weighted-$\ell_1$ optimization of Walsh coefficients and a sparse recurrence requiring $O(\sqrt{m})$ terms on balanced rectangles, avoiding $O(m^4)$ dense constraint storage. Across 160 geometries, weighted-$\ell_1$ completion reduces controlled-NOT (CX) counts relative to four alternative extensions in all 96 cases with unused binary codes under Gray-code synthesis. On an independently specified 60-case cohort, median reductions relative to virtual-coordinate extension are 28.0%, 53.9%, and 21.6% at six, nine, and twelve sites. Under generic diagonal synthesis, reductions decrease to 10.9%, 1.3%, and 0.7%, demonstrating compiler dependence. Additional ancillas reduce mixer serialization, but token circuits remain deeper than one-hot baselines. Ideal placement simulations show baseline-dependent solution quality, with classical search performing better. OpenROAD integration takes 72 QAOA and 216 classical placements across six RTL designs through clock-tree synthesis and global routing with zero overflow. Completion improves phase construction; no end-to-end advantage is established.

Recent LLM-based approaches to control either invoke a language model to select actions or synthesize world models that require planning at every decision, introducing latency that can limit real-time use. We introduce Code to Control, an approach that synthesizes Python controllers which execute directly as policies. Code to Control separates program structure from parameters. An LLM synthesizes the controller structure, while derivative-free search fits its parameters for continuous control using feedback from the environment. Once learned, the resulting controllers require neither LLM inference nor planning at decision time, enabling real-time gameplay and, under our timing protocol, faster action selection than a PPO policy. Across a suite of Atari games, Flappy Bird, and MuJoCo tasks, Code to Control outperforms planning-based program synthesis methods, remains competitive with deep reinforcement learning while using fewer environment interactions, transfers across substantial changes in environment dynamics, and scales to complex locomotion tasks.

We study noncooperative multi-vehicle games among heterogeneous autonomous vehicles, where each vehicle adopts a decentralized closed-loop policy based on its own state, and optimizes an objective that depends on other vehicles through potentially asymmetric interaction weights. We develop an $\alpha$-potential game framework that reduces the computation of an approximate Nash equilibrium (NE) to the minimization of a single auxiliary $\alpha$-potential function. We explicitly construct this $\alpha$-potential, establish the existence of its minimizers, and characterize the equilibrium approximation error $\alpha$ in terms of interaction asymmetry. We further introduce vehicle-specific scaling to reduce the effective interaction asymmetry, thereby tightening the equilibrium approximation and, in important cases, recovering an exact NE despite asymmetric interactions. We also derive social-efficiency guarantees for the potential-selected policies, revealing how the interaction structure shapes worst-case efficiency. Numerical experiments demonstrate the flexibility of the framework in capturing heterogeneous vehicle interactions, collision and obstacle avoidance, lane changing and overtaking under different traffic configurations, and priority-based intersection crossing.

Flapping-wing vehicles regulate aerodynamic forces and moments through coordinated variations in stroke, deviation, and pitch motion. Because the resulting loads depend on both the instantaneous wing configuration and its preceding motion history, recovering suitable wing kinematics from a desired aerodynamic trajectory is a challenging inverse problem. Existing sequence models capture temporal dependencies, while spectral methods can exploit the periodic structure of flapping motion. However, unrestricted frequency-domain augmentation may interfere with temporal representations and produce inconsistent corrections across kinematic variables and prediction horizons. We propose the Bounded Channel-Adaptive Spectral Residual Gated Recurrent Unit (BCS-GRU), which retains recurrent temporal prediction as its primary representation and restricts spectral information to a controlled output-specific correction. We further introduce a causally aligned forward-consistency objective that evaluates predicted kinematics through a separately trained and frozen aerodynamic surrogate. Experiments under a unified episode-level protocol show that BCS-GRU improves inverse prediction over recurrent and adaptive-spectral baselines, with greater benefits at longer prediction horizons. Forward-consistent fine-tuning further improves surrogate-based aerodynamic consistency while maintaining mean kinematic accuracy. These results demonstrate that controlled spectral correction and forward-consistent learning provide an effective framework for history-aware inverse modelling of flapping-wing aerodynamics.

Pet bereavement often receives little social recognition. Generative AI can give a continuing bond a responsive voice, but a comforting reply may also claim authority to forgive or request attention. We investigate how a memorial can support connection without turning remembrance into obligation. Through Research through Design, we developed Afterglow, a mobile world connecting private remembrance, human witnessing, and symbolic Pet messages. A formative survey (N=57) informed the initial design, followed by six online roundtable walkthroughs with 20 unique participants across two prototype iterations. Our interpretive analysis develops tensions between returnable connection and emotional obligation, recognizable likeness and ontological clarity, and protective intervention and surveillant authority. We contribute Legible Restraint, a cross-layer requirement that limits on relational authority survive changes in speaker, generation context, trigger logic, data use, and participation. Its temporal consequence, Designing for Goodbye, keeps remembrance available without making continued use a condition of care.

In this paper we give a polynomial-time reduction from the Subset Sum Problem to unconstrained static output feedback stabilization, establishing its NP-hardness. The construction uses two scalar building blocks to impose approximate discrete choices and a weighted-sum constraint. Once the problem is encoded, we relate the stability of the $N+1$ decoupled loops encoding the problem to that of a coupled plant. This is accomplished by leveraging frequency separation via a band-pass transformation, and comparing the root counts of the true characteristic polynomial with that of the different blocks in different regions of the right half-plane. The construction then naturally bounds every stabilizing gain and gives explicit modulus-margin estimates, ensuring that the plant data have polynomial binary encoding length.

In graphical two-choice allocation, each arriving ball is assigned to one endpoint of a random edge. We study rules whose decision is a monotone function of the two endpoint loads, allowing edge-dependent thresholds and fresh randomization. Such a rule has an exact unit-discrepancy coupling: adding one ball to the initial state produces one tagged discrepancy at every later time. We represent the tag by conditional-expectation projections on the marked edge space and obtain diffusive displacement bounds. A transport-volume inequality then converts slow propagation of influence into lower bounds for the load gap. On the cycle with $n$ vertices, from every initial distribution and at every physical time $t\ge 1/n$, the expected gap is at least a constant times $\min\{\sqrt n,t^{1/4}\}$, and the gap exceeds this scale with probability at least $1/8$. After exactly $k\ge1$ allocations, the corresponding scale is $\min\{\sqrt n,(k/n)^{1/4}\}$. No stationarity, symmetry, recurrence, or moment assumption is used. A smoothed threshold rule in the same class has expected gap $O(\sqrt n\log n)$ up to any fixed polynomial time horizon, so the saturated cycle bound is sharp within the class up to a logarithmic factor. The general inequality also yields a lower bound of order $\sqrt{L/K}$ on the $L\times K$ rectangular torus $C_L\square C_K$; combined with a strategy-independent logarithmic bound, this gives order $\sqrt{L/K}+\log(LK)$. These results separate endpoint-local rules from global-information strategies that achieve polylogarithmic gaps on cycles. All numbered results are verified in Lean.

We study finite-blocklength lossy transmission of a Gaussian memoryless source over memoryless, possibly nonstationary block erasure channels under an excess mean-squared distortion criterion. We derive computable nonasymptotic achievability and converse bounds and establish matching third-order, channel-uniform normal approximations. Our achievability bound exactly evaluates the ensemble-average excess-distortion probability of a specified random coding scheme and improves corresponding specializations of known general one-shot bounds. Our converse conditions on the erasure pattern and source energy and combines spherical-cap and volume bounds, with the spherical-cap term providing the geometric prefactor needed for the matching third-order term. In our channel-uniform normal approximation, we show that for fixed distortion ratio, target excess-distortion probability, and block size, the sufficient and necessary information-balance conditions have the same dispersion and $\frac{1}{2}\log k$ terms, where $k$ is the source blocklength, and differ only by bounded remainders that are uniform over the channel blocklength and erasure-probability profile. Numerical evaluations compare the nonasymptotic JSCC bounds and their common third-order normal approximation with an optimized symmetrized SSCC achievability benchmark.

Coding agents have become real users of high-performance computing (HPC) systems, yet today's HPC abstractions, interfaces, and policies remain designed for human-driven workflows. In our measurement, users running coding agents are only 19.5% of the observed population, but account for 55.8% of job submissions, 29.1% of CPU core-hours, and 42.7% of GPU-hours. Agents are not simply faster humans. They issue commands at 20.8x the human rate, decompose work into fine-grained explore-modify-execute loops, and pursue open-ended goals through trial-and-error campaigns that continue through nights and weekends. These behaviors strain abstractions built for human timescales, surfacing as control-plane pressure on the scheduler, metadata-intensive I/O on bandwidth-provisioned filesystems, repeated rediscovery of what earlier sessions already learned, and new prompt-injection and policy-enforcement surfaces. Neither banning agents nor treating them as ordinary users is sustainable. We instead argue for co-design, that facilities should treat agents as first-class principals where agents become facility-aware tenants. We outline the resulting challenges and opportunities in compute, storage, agent memory, and safety.

LLM watermarking has become an effective approach to distinguishing AI-generated text from human-written text by embedding detectable patterns during generation. However, a small post-generation edit may change the meaning of the text without removing its overall watermark signal, creating a risk that the modified content is still attributed to the original model. We propose Anchor-ECC, which incorporates the error-correcting code (ECC) constraints and explicit boundary anchors into the watermark structure and pairs them with a dynamic-programming decoder to detect and localize post-generation edits. Across Qwen3-8B, Mistral-7B-Instruct-v0.3, and OPT-125M, the approximate-hard setting achieves about 99.7% block-level true positive rate (TPR) with at most 7.6% false alarm rate (FAR) for edit detection under mixed insertions, deletions, and substitutions, while preserving the distinction between watermarked outputs and unwatermarked text. Additional quality experiments identify lower-perplexity configurations that retain strong edit-detection performance. Together, these results extend LLM watermarking from source identification to local integrity verification while supporting configurable trade-offs between detection reliability and generation quality.

Modern multimodal models bring generation and understanding into a single unified system, which enables them to provide and learn from their own feedback. Motivated by this unified capacity, we introduce UniEvo-VL, a self-evolving framework for multimodal models to learn from this constructive self-correction feedback during test-time compute. Instead of relying on a separate, often larger, teacher, we leverage their self-critiques as privileged information and ask a single multimodal model to act as both teacher and student with different contexts. The student only sees the vanilla question, while the teacher conditions on the privileged critique. Then training minimizes the per-state divergence between their denoising diffusion distributions over the student's own sampling trajectories. Experiments demonstrate that UniEvo-VL improves the image generation capabilities of multimodal models, while maintaining their sensitivity to additional reflection information. Specifically, we build on top of the open-source Qwen-image-2512 and observe a significant performance gain from 0.747 to 0.808 on GenEval and from 32.97 to 35.53 on GenEval2 Soft-TIFA. Moreover, attempts with more powerful external critics (e.g., GPT5.6-Luna) show that multimodal models with strong judge capabilities can anticipate a higher self-evolving ceiling. Last but not least, mixed text-rendering outcomes show that our self-improvements may not be uniform across different tasks. Our study aims to shed light on the current hot recursive self-improvement research line to enhance the user experience when using multimodal models without external supervision or guidance.

Readout-side residual hybrids concatenate raw inputs with measured quantum features. Under single-example gradient sharing, a biased first linear layer admits standard analytic recovery of its input, so the bypass exposes raw coordinates without requiring inversion of the quantum circuit. We audit this mechanism using two tabular datasets, four architectures, and metrics conditioned on feature visibility. Iterative gradient matching gives median full-record PSNR of 73-96 dB for residual and input-only heads. Quantum-only heads score 8-11 dB on the full record but 54-96 dB on the six input coordinates they actually encode. These are reconstruction results for the tested six-input, six-observable circuits, not a general statement about quantum encodings. A loss-threshold membership attack remains near chance. The contribution is a visibility-conditioned privacy audit: omitted coordinates must not be credited as protection supplied by quantum processing, and near-exact PSNR differences must not be interpreted as meaningful privacy rankings. Our findings concern individual gradients and do not establish leakage under aggregation or multiple local training steps.

Experimental investigation and modeling of flapping-wing aerial vehicles are limited by the scarcity of large-scale records that temporally align wing kinematics, aerodynamic responses, and flight states. Existing datasets are often limited in scale and affected by measurement noise and temporal misalignment between rapidly varying wing motion and the associated dynamic response, particularly during high-frequency flapping. We introduce FlapKAD, an episode-structured simulation dataset comprising 2,000 rigid-wing flight episodes and 720,152 valid time steps, with bilateral wing kinematics, aerodynamic responses, and flight states recorded synchronously within a common clock. FlapKAD supports a unified bidirectional sequence-prediction benchmark constructed from the same temporally aligned episodes. The forward task predicts future vertical force coefficients and body vertical velocity from histories of realized flap and twist angles, whereas the inverse task reconstructs future flap- and twist-angle trajectories from the corresponding response histories. A benchmark of eight representative time-series architectures across multiple prediction horizons reveals direction- and horizon-dependent model behavior, systematically higher reconstruction errors for twist angle than for flap angle, and no consistent advantage from increased architectural complexity. FlapKAD provides a reproducible dataset and benchmark for studying coupled wing-kinematic, aerodynamic, and flight-state dynamics in flapping-wing aerial vehicles.

Steering large language models typically relies on linear, context-independent interventions in activation space, an assumption that recent work has challenged and that can induce an information bottleneck when a fixed representation must encode many behavioral distinctions. We introduce MetaSteer, a method that learns nonlinear interventions with context-dependent effects and applies them to attention projection matrices, producing activation effects that vary with the input context by construction and requiring no linear concept-geometry assumption. Framed as preference-based optimization, MetaSteer is trained once on a pooled preference corpus and transferred zero-shot to unseen concepts and out-of-distribution contexts. We find that, despite using low-rank adapters, MetaSteer induces structured, context-dependent changes in hidden-state trajectories while partially preserving aspects of their local trajectory dynamics, including velocity and curvature. We evaluate MetaSteer on three controlled text-generation benchmarks and three agentic settings across multiple model families and scales. MetaSteer matches or outperforms strong task-specific steering baselines on most aggregate comparisons in the zero-shot regime. Across the evaluated settings, stronger text-generation steering is associated with stronger agentic steering performance. We further discuss geometric trajectory effects, capability retention, and safety considerations raised by transferable steering.

Large Vision-Language Models (LVLMs) commonly perform spatial reasoning through chain-of-thought (CoT), encoding intermediate reasoning as autoregressive sequences of discrete language tokens. Such hard thinking requires committing to a single token at each step, even when the correct spatial interpretation remains uncertain. This early commitment constitutes premature discretization: an incorrect token selection can propagate errors through subsequent reasoning. We propose Soft Spatial Reasoning, a post-training framework that introduces soft thinking for spatial tasks in LVLMs. At each intermediate reasoning step, the LVLM forms a continuous soft state by mixing token embeddings rather than selecting a single token, allowing multiple candidate continuations to influence the next step. The appropriate degree of softness, however, can vary across reasoning steps: retaining multiple candidates may preserve a useful spatial interpretation, but if those candidates imply conflicting spatial relations, mixing them may interfere with subsequent reasoning. At the core of Soft Spatial Reasoning is AdaptSoft, a controller that uses the current hidden state and predictive uncertainty to adapt the degree of softness at each reasoning step. To train AdaptSoft, we introduce a gradient-alignment learning objective that provides a step-specific learning signal for softness control without intermediate reasoning supervision. Across diverse spatial benchmarks, Soft Spatial Reasoning outperforms hard and fixed-soft CoT baselines using the same backbone, as well as a range of existing LVLMs. The source code is available at https://github.com/rafiibnsultan/Soft_Spatial_Reasoning

Large Vision-Language Models (LVLMs) have made remarkable progress across visual perception tasks, yet spatial reasoning remains a persistent weakness, especially for questions that require reasoning over visual space. Recent spatial-reasoning methods incorporate generated grounding, where models predict bounding boxes, masks, or other localization outputs for task-relevant objects as part of their reasoning trace. However, these approaches typically optimize final-answer correctness alone, allowing correct answers to be rewarded even when the model does not reason from confidently localized task-relevant objects. We introduce SpatialCORE (Spatially COnfident REasoning), a post-training framework that turns the model's own confidence in generated grounding into a learning signal for spatial reasoning. Its central idea is to reinforce grounding that is both accurate and confident, encouraging the model to reason from confidently localized task-relevant objects. SpatialCORE realizes this through a self-regulating spatial reward that weights each predicted bounding box's matching quality by its coordinate-token confidence. An answer gate further ties grounding optimization to final-answer correctness. SpatialCORE achieves state-of-the-art results among open-source and specialized spatial reasoning models across diverse benchmarks, and transfers effectively in zero-shot settings to unseen data distributions. The source code is available at https://github.com/rafiibnsultan/SpatialCORE.

This paper proposes a control strategy for path following that is model-free and relies solely on deceleration for skid-steering planetary rovers navigating deformable loose terrain under continuously imposed traction asymmetry. While conventional controllers that are based on kinematics frequently cause slip-sinkage entrapment by accelerating the wheels during path correction, the proposed approach prevents this failure by setting an upper limit on the maximum commanded velocity. Heading correction is achieved solely through the selective deceleration of the outer wheels, which are located on the outside of the turn, driving them into a negative slip regime to act as a mechanical anchor. The system was evaluated using the four-wheel independent-drive rover EX1 under an asymmetric wheel configuration with different left and right grouser heights that induces significant deviations from the path. Experimental results demonstrate that this deceleration-only control successfully suppresses accumulated lateral drift across various velocity regimes up to 0.7 m/s without causing sinkage. Crucially, direct force measurements from onboard multi-axis sensors provide important empirical evidence of the underlying terramechanics, proving that the targeted deceleration establishes dynamic load equalization across the chassis and completely restores the native thrust capability of the opposite driving wheel.

We describe our winning entry to the Waymo Open Dataset 2D Video Panoptic Segmentation Challenge. The task asks for a semantic class at every pixel of every frame and, for countable objects, an identity that holds across 100 frames and across five overlapping cameras. We build on DVIS++, a cascade of a segmenter, a tracker, and a refiner, as our baseline. We propose hard region supervision (HRS) to improve the baseline. In particular, we use the baseline to define the hard region as where it makes mistakes, and design a loss and an auxiliary prediction head for this region. The auxiliary head is used only in training and removed at test time, so at inference the model trained with HRS has the same architecture as the baseline. In addition, we propose three test-time steps that further improve the results: a two-model ensemble, a merge of the segmenter's output into the final panoptic map, and cross-camera identity linking. On the challenge test set, our entry reaches 0.3547 wSTQ, 0.2071 wAQ, and 0.6075 mIoU, ranking first on all three metrics. It is 3.6 wSTQ points ahead of the second entry and 2.4 points ahead of our DVIS++ baseline.

Long-horizon agentic workflows require models to sustain repeated state-dependent actions all while the context grows, sub-task complexity changes, and new data arrives. Each situation represents an independent axis along which an agent may fail. An agent reconciling a long ledger, for example, must repeatedly read its state, update the correct record, and preserve alignment across thousands of outputs. A model may accept the entire ledger yet lose its place or stop applying the operation consistently as generation proceeds. We introduce Long-Transduction, a controlled diagnostic that tests a model's ability to stay on task during long generation while continuously reading, mutating, and outputting input-context dependent operations such as arithmetic, sorting, variable lookups, and table transformations. Long-Transduction evaluation independently varies local task complexity, input data formatting, and context length isolate failures along each axis. We evaluate seven open-weight models, finding a 62.8% decrease when scaling context length from 4-128K, a 36.5% decrease when varying input format, and a 39.9% decrease by increasing local task complexity. Together, these failures represent critical liabilities in long-horizon agentic workflows.

We study the problem of sampling from a continuous density $\pi\propto \exp(-V)$ on $\mathbb R^d$, where $V\in C^2(\mathbb R^d)$ has a $\beta$-Lipschitz gradient and $\pi$ satisfies a logarithmic Sobolev inequality with constant $\alpha^{-1}$, and write $\kappa = \beta/\alpha$. We introduce the Gaussian cloud sampler, which achieves total variation accuracy $\varepsilon$ using $\widetilde O(\kappa d^{1/5}\,\text{polylog}(1/\varepsilon))$ gradient queries in expectation. The algorithm uses first-order rejection sampling (FORS) to correct the law of smoothed Picard HMC trajectories. To do so, we represent the iterates of the ideal Picard iteration via Gaussian clouds, whose centers are never evaluated, and we develop a suite of likelihood correction gadgets which only use samples from this indirect cloud representation.

Humanoids are increasingly capable of tracking complex whole-body motions, but physical interaction introduces a different challenge. When a robot makes contact with a person or the environment, it needs to respond to external forces while preserving the motion needed for the task. This response can vary across directions in the end-effectors and on the body. For example, an end effector may need to accommodate contact force in one direction while maintaining motion accuracy in another, while the robot body may resist an external force or move with it. We present a compliance framework for humanoid loco-manipulation that combines directional and tunable end-effector (EE) compliance with selectable root compliance for external force rejection or force following. A hierarchical reinforcement learning controller modulates a fixed whole-body tracking policy through high-level EE and root commands, while interaction forces are estimated from proprioceptive history. Our simulation and real-world experiments on a humanoid demonstrate directional stiffness control, online stiffness adjustment, distinct root compliance, compliant manipulation, and collaborative carrying.

We develop a task-relative framework for certifying residual wrench authority after hover and contact loading in multirotors with bounded actuators. Using convex geometry, we derive signed margins for prescribed convex reserves, including Euclidean balls and weighted ellipsoids. We obtain computable reserve certificates from actuator-interiority bounds to support slack-maximizing allocation. To preserve the required reserve, we propose command projection onto a tightened feasible set. We then connect the available reserve to structured gain synthesis and certify a local tracking region that respects actuator limits. Within this region, we establish nominal exponential convergence and robust ultimate boundedness. For the prescribed task and morphology family, we show that the optimized octarotor attains a larger margin than the optimized hexarotor at equal total thrust. We evaluate the proposed framework in closed-loop simulations of sustained rigid-wall contact using a fixed-geometry octarotor and a variable-tilt quadrotor with matched installed thrust. The quadrotor's local certificate admits a higher normalized push limit for the prescribed task family. In tests beyond the realizability boundaries, we observe the predicted loss of authority through rotor-thrust limits for the octarotor and tilt-servo limits for the quadrotor.

Production LLM serving stacks combine an inference engine's local prefix cache with a shared KV-cache tier for fleet-wide reuse. The local cache distinguishes requests by adapter, weight configuration and sharing domain, but the shared tier may key entries only by token content and coarse model metadata. This boundary erases provenance and lets identical tokens under incompatible computational or sharing contexts collide. We call this composition gap provenance-blind reuse and present its first systematic study. A source audit of three vLLM connectors confirms the structural omission, while runtime experiments reproduce it across vLLM and two SGLang releases, 12 models from 7 families (0.5 B-32 B), and over 160 configurations. Cross-adapter collisions reduce accuracy from 0.94 to 0.64, incompatible KV representations reduce reasoning accuracy to zero, and salt omission enables 93% prompt identification from timing. We formalize the missing guarantee as the KV provenance contract: for a declared dimension registry, shared keys must be injective over computational and sharing provenance, with identities stable across workers. Any dimension with a stable identity can therefore be added without connector-specific key logic. A canonical descriptor binds per-request and per-worker provenance into lookup and store keys, while a differential checker detects dimensions that change KV state without changing the key. Implemented in vLLM and SGLang 0.5.20 across three cache paths, provenance binding eliminates unsafe reuse while preserving legitimate sharing. Hit-path latency changes remain within 0.34 ms and below run-to-run variation; retention grows with provenance diversity.

Foundation models for computational pathology are usually evaluated using AUC and accuracy, while calibration is often left untested. This matters because a model can be accurate on average but still assign overly confident probabilities to cases that are difficult even for pathologists. We study calibration across eight pathology foundation models. Using pathologist agreement as a measure of diagnostic difficulty, we find that calibration error is consistently higher on low-agreement cases than on high-agreement cases. This pattern is not apparent from aggregate expected calibration error (ECE) alone. We then propose synthetic agreement calibration, a method for improving calibration without collecting multi-annotator labels. Given a trained linear probe, we select high-confidence embeddings as class anchors and interpolate between anchors from opposite classes. The interpolation weights encode a continuous notion of diagnostic ambiguity, which we use as a synthetic agreement signal to retrain the probe with agreement-aware label smoothing. On MHIST, which includes annotations from seven pathologists, synthetic agreement calibration recovers most of the calibration improvement obtained by label smoothing based on real pathologist agreement, while substantially reducing low-agreement ECE relative to the uncalibrated baseline. Discrimination metrics are preserved. On PatchCamelyon and BreakHis, public histopathology datasets without multi-annotator labels, the method improves calibration across the evaluated foundation models, whereas annotator-dependent approaches cannot be used without additional expert annotation.

Recent works have demonstrated that quantum oracles have subtle behavior, as access to inverse, conjugate or controlled queries can exponentially change the query complexity of certain tasks. Inspired by these works, we introduce the notion of meta-complexity of quantum relativization. We ask: for any two quantum complexity classes, under which "types" of quantum oracles are they equal or separated? Different pantheons of oracles (or quantum oracle types, e.g. unitary vs state, poly- vs superpoly-dimensional, closed under inverse or not) form a partially ordered set based on their power in separating complexity classes. Moreover, two oracle pantheons A and B are separated if there exists a pair of complexity classes that are separated under an oracle from pantheon A but yet the complexity classes are equivalent under all oracles from pantheon B. We show that this meta-complexity can be nontrivial by giving a complete classification, within the family of oracle pantheons defined in this paper, of which models can separate the complexity classes $\mathsf{PostBQP}$ and $\mathsf{PreciseBQP}$, the exponentially precise variant of $\mathsf{BQP}$. Both classes equal $\mathsf{PP}$ in the unrelativized setting. Within our taxonomy, they remain equal relative to real or polynomial-dimensional unitary oracles and whenever inverse or conjugate access is supplied. In contrast, we give a separation relative to forward-only complex diagonal unitaries of superpolynomial dimension, as well as a separation relative to single-qubit state-preparation oracles. We view this as a test case for the meta-complexity of oracles which underscores the subtlety inherent to the relativization of quantum complexity classes.

A cumulative token cap can fall inside a mathematical derivation, forcing a test-time controller to choose between stopping at the cap (strict) and allowing the current attempt to finish (advisory). We measure this boundary choice with paired offline replays of 19,200 public traces: 120 AIME, BrUMO and HMMT problems and two archive configurations of one model. Candidate order and a 16-attempt cap are fixed, and answer selection is blind to reference answers and correctness labels. Three findings emerge. First, at the 4k cap, most advisory accuracy gains replace abstention with a correct answer; strict stopping pays for an unfinished prefix that the completed-only selector cannot use. Second, comparisons along realized cost differ from same-cap comparisons: advisory 4k in low has higher accuracy than strict 8k at comparable mean completion cost, while in high its observed accuracy is 0.42 points below strict 32k using 59% of its mean tokens. These aggregate comparisons do not establish equal-compute superiority or accuracy equivalence. Third, increased candidate coverage does not guarantee higher answer accuracy: a log-probability selector loses accuracy while coverage rises, including after a source-grade consistency repair. Same-cap majority-accuracy differences shrink below 1.3 percentage points at 32k. Budget curves should jointly state the cap, realized cost, eligible candidates, stopping rule and selector information.