Papers

Latent LiDAR pipelines suffer from flying pixels: convolutional VAEs blur sharp radial depth discontinuities, yielding edge depths that back-project to points floating between surfaces. We identify this as a major, directly correctable decoder bottleneck and introduce CRISP: a pixel-space diffusion decoder with a backbone-agnostic latent adapter, DiT-based denoiser, and support mask predictor. CRISP replaces video-VAE and LiDAR-native decoders alike while keeping the encoder and latent generator fixed. Across KITTI-360, SemanticKITTI, and nuScenes, replacing only the decoder reduces FSVD/FPVD by 50.5% on average across frozen backbones; for generic video VAEs, the reductions reach 71%/74%. On the LiDAR-native LiDM backbone, FRID drops by 71%, with the largest gains at depth discontinuities. In a pretrained LiDM world model, the same zero-shot replacement improves FSVD by 15.5%, narrowing the sim-to-real gap. Code and checkpoints will be released at https://andrea25512.github.io/CRISP/.

Personalized conversational shopping requires maintaining preference consistency over multi-turn interactions, where users reveal constraints gradually. Existing approaches often rely on static profiles and do not explicitly control long-horizon interaction behavior. We propose a multi-agent, multimodal Retrieval-Augmented Generation (RAG) framework that decomposes dialogue state tracking, recommendation retrieval, preference-aware reasoning, and response generation, while integrating product metadata, product reviews, image-derived descriptions, and user historical reviews. To evaluate interaction-level quality, we adopt a trajectory-level protocol with four dimensions: Global Preference Consistency, Cumulative Information Synthesis, Interaction Trajectory, and Tone Consistency. On an Amazon Reviews 2023 benchmark, retrieval-enabled variants outperform a no-RAG baseline on automatic trajectory metrics (average 4.82 vs. 3.74). In a small real-user study ($n{=}5$), the Full variant achieves the highest mean overall rating (4.60 vs. 2.20 for Baseline), providing exploratory evidence that role decomposition plus user-centric retrieval improves perceived personalization.\footnote{Code and dataset are available at: https://github.com/RenaGao/Multimodel_RAG_Indexing

CLIP serves as a foundational vision-language model and the de facto vision encoder for downstream VLMs such as LLaVA. Post-training offers a lightweight route to refine CLIP, but recent work argues that the standard contrastive loss is unsuitable for post-training due to catastrophic forgetting under small batches, motivating designs that abandon the contrastive objective in favor of distillation. We revisit this premise and find that, for the InfoNCE objective, the reported forgetting is driven primarily not by insufficient negatives but by an inappropriate magnitude of the contrastive temperature $\tau$: with $\tau$ set sufficiently small, contrastive post-training improves rather than degrades the pretrained CLIP, which we explain through the temperature dependence of the InfoNCE gradient. Building on this finding, we propose ComCLIP, a lightweight single-epoch post-training recipe that freezes CLIP's text encoder—so the refined vision encoder is a drop-in replacement with unchanged architecture and inference cost—and trains the vision encoder with a properly-tempered contrastive loss, an MSE anchoring loss against the original CLIP, and a relational distillation loss from DINOv2. Over multiple seeds, ComCLIP matches the self-distillation baseline CLIP-Refine on zero-shot classification while significantly improving the transferability of visual features, measured by linear probing ($48.99$ vs.\ $42.28$ on ViT-B/16), and on ViT-L/14 it also improves MMVP over CLIP-Refine ($24.20$ vs.\ $19.01$); CLIP-Refine remains stronger on image-text retrieval. Used as a drop-in vision encoder for LLaVA-1.5-7B without re-aligning the projector or LLM, ComCLIP yields no net change across $8$ VLM benchmarks, i.e., the refinement does not break downstream compatibility. Code and models are available at https://github.com/showstarpro/ComCLIP.git.

Continual learning remains challenging for large language models, which must enable models to acquire new skills and knowledge without degrading existing capabilities. Existing approaches typically address this challenge by carefully designing how model parameters are updated. In contrast, prompt optimization avoids costly parameter updates while achieving competitive or even superior performance to reinforcement learning methods such as GRPO on individual knowledge-intensive and reasoning tasks. This raises a natural question: Can prompt optimization, as an efficient adaptation approach, be directly applied to continual learning? Our analysis shows that, under sequential task adaptation, it suffers from catastrophic forgetting, while optimized prompts accumulate rules that overfit to local task distributions. To address these limitations, we propose Evolving Functional REpertoires (EFRE), which replaces a single prompt with a repertoire of functions that evolves as new tasks arrive: compatible updates refine existing functions, while conflicting updates trigger the emergence of new ones. On a three-task continual-learning stream, EFRE achieves a final average performance 7.50 percentage points higher than GRPO. Moreover, after adaptation to the Bio task, its performance on FinQA decreases by only 1.56 percentage points, compared with 25.10 percentage points for the base prompt optimization method. We further instantiate EFRE in a minimal agent system and observe consistent improvements across different backbone models. Overall, these results demonstrate EFRE's strong performance in continual learning for large language models and highlight its substantial potential for continual learning in advanced agent systems.

Gate insertion (GI) quantum error mitigation replaces involutory gates such as CZ and CX with 2j+1 repetitions of the same gates to nominally amplify noise by the same factor. Analytical and experimental results show that this amplification scaling fails for non-commuting gate noise, as typically encountered in realistic devices. Here we introduce Heisenberg-echo symmetrization (HES), which combines standard GI with an echo inspired by the Heisenberg representation to cancel the leading GI error. In our simulations, HES reduces the absolute fidelity error by one to two orders of magnitude compared with GI. Unlike the more accurate Layered-KIK method, HES is compatible with any hardware and requires only circuit-level access. It enables virtual noise scaling across platforms, potentially reducing sampling overhead substantially. Its agnostic noise amplification also provides full resilience to temporal noise drift. For pure initial states, however, high-rank observables can substantially increase the echo variance relative to non-echoed circuits, whereas rank-one observables incur no additional sampling overhead compared with GI or Layered-KIK. HES is therefore particularly suited to rank-one or nearly rank-one observables.

Contemporary decoder-only large language models (LLMs) have demonstrated strong capabilities across a wide range of domains. However, existing pretraining paradigms for gloss-free sign language translation (SLT) are largely designed around conventional encoder-decoder pretrained language models, which limits their direct applicability to decoder-only LLMs. To address this limitation, we propose SignRAG, a unified framework combining hierarchical pretraining, target-domain retrieval augmentation, and retrieval-aware reinforcement fine-tuning. Hierarchical pretraining first learns linguistically grounded sign representations and then jointly aligns the sign encoder with an LLM, mitigating cross-modal optimization imbalance. For downstream adaptation, SignRAG complements parameter-based fine-tuning with a target-domain retrieval gallery that provides instance-specific translation cues. To ensure that retrieved contexts are used appropriately, we further introduce Retrieval Utility-Guided Reinforcement Fine-Tuning (RUG-RFT), which combines translation-quality and retrieval-utility rewards to encourage beneficial retrieval use while suppressing harmful reliance. Experiments on multiple SLT benchmarks establish new state-of-the-art performance. In particular, to the best of our knowledge, SignRAG is the first gloss-free approach to outperform gloss-supervised methods across all reported metrics on CSL-Daily. Our code has been released at \href{https://github.com/shahelaojieraozhi/SignRAG}{GitHub}, together with models of different sizes to support future academic research.

Retrieval-augmented generation (RAG) grounds language models in external corpora. Agentic RAG enables iterative search, yet exposes the model to isolated chunks without document structure, making it difficult to distinguish relevant evidence from chunks that merely resemble the query. Structure-aware methods such as PageIndex navigate using the structure but cannot scale to structures of large corpora that cannot fit in the LLM context. Hence, they first commit to a single document using a document retriever and cannot recover from a wrong choice. We propose RIT-RAG (Retrieval-Induced Tree RAG), which combines content retrieval with structural navigation. Offline, RIT-RAG builds a tree for each document from its table of contents or sitemap. At query time, it retrieves a broad set of chunks and uses their positions to induce manageable sub-trees, potentially across multiple documents. An LLM agent navigate these sub-trees, selectively reads promising nodes, and reformulates queries when needed. Thus, retrieval proposes where to look, while the agent decides what to read. Across financial, scientific, and customer-support benchmarks, RIT-RAG achieves the highest answer accuracy among vanilla, graph-based, and agentic baselines. On EntQABench, our new benchmark of 2.84 million technical-documentation webpages, it improves accuracy by 6.6–11.4 points over the strongest baseline across three LLMs.

We consider entire solutions $u$: $\mathbb{R}^n \rightarrow \mathbb{R}$, $n \ge 2$, of the equation $\operatorname{div}(\nabla f(\nabla u)) = 0$ with convex density $f$: $\mathbb{R}^n \rightarrow \mathbb{R}$ given for example by $f(\nabla u) = \sum_{i=1}^n |\partial_i u|^{p_i}$ with exponents $p_i \ge 2$ including Giaquinta's example for the choice $p_1 = \dots = p_{n-1} = 2$, $p_n = 4$. If $n \le 2 + \min\{p_i\}$ holds, then we have the Liouville property: the boundedness of $u$ implies its constancy.

Muography with scintillator-based detectors enables the investigation of density distributions within large structures. The Hankuk Atmospheric-muon Wide Landscaping (HAWL) project is the first real-time portable muon tomography. It successfully charted the mountainous region of eastern Korea by measuring cosmic-ray muons with a detector operated in a moving vehicle. HAWL achieved a tunnel length accuracy of 6.0 % and an overburden measurement range of 8-400 meter-water-equivalent (m.w.e.). The HAWL upgrade (HAWLU) features a more compact design and enables trajectory reconstruction. Using this upgraded system, we measured tidal water-level variations with meter-scale precision over 31.6 days in the Boryeong Undersea Tunnel. The measured muon rate decreased by $1.72 \pm 0.04$ % per meter of increasing tidal level, while the reconstruction of undersea tunnel scan independently validated the detector performance. These results demonstrate the capabilities of dynamic muography in two approaches, one with a portable detector in motion and the other with a detector measuring a target that changes over time.

The role of adaptive coupling in controlling collective noise-induced dynamics is demonstrated on an example of stochastic resonance in an ensemble of coupled overdamped bistable oscillators. Specifically, it is found that tuning the adaptive coupling parameters allows one to either enhance or suppress stochastic resonance and to shift the optimal noise intensity that yields the most regular stochastic oscillations. All the revealed effects are studied in numerical simulations for two kinds of adaptive coupling models. The first one is a simplified, phenomenological model designed such that the coupling becomes stronger for oscillators with larger differences between their dynamical variables (two configurations for local interaction and one option for global coupling are under consideration). The second one is a local memristive coupling model that provides a physically motivated realization of adaptive interactions with similar properties. In all the considered cases, adaptive coupling is shown to be a flexible stochastic resonance control tool, both in the absence and in the presence of the static coupling component.

Spatial coding agents significantly improve spatial reasoning in Multimodal Large Language Models (MLLMs) by using external tools to generate verified execution traces. However, this paradigm inherently suffers from prohibitive inference-time overhead and external dependencies. In this paper, we explore whether an MLLM can internalize this agentic capability to operate entirely tool-free. We begin with a simple observation: prompting an MLLM with summarized execution traces of a spatial coding agent naturally unlocks the model's internal spatial Chain-of-Thought (CoT). Motivated by this, we introduce SpatialOPSD, an on-policy self-distillation framework that internalizes spatial reasoning into a standalone MLLM by formulating verified agent traces as privileged information. To mitigate privileged-information leakage during distillation, we introduce Repetition-Aware Distillation, which combines repetition masking with unlikelihood regularization. Experiments across multiple benchmarks demonstrate that self-distilling SpatialOPSD achieves higher average accuracy than SFT and GRPO on both spatial and OOD datasets, exhibiting superior performance and generalization.

In this paper, we propose a relaxed CQ algorithm with dynamic stepsizes for the split feasibility problem (SFP), where the relaxed parameters are chosen adaptively. We establish the weak convergence of the proposed algorithm under standard assumptions and further derive its linear convergence under the bounded linear regularity property of the SFP. Numerical results on LASSO problem demonstrate that the adaptively over-relaxed CQ algorithm with dynamic stepsizes outperforms the dynamic CQ algorithm proposed in [10].

A trained 3D Gaussian Splatting (3DGS) scene ships as a portable parameter array that can be copied, pruned, requantized, or repackaged outside its training pipeline, so ownership evidence is most useful when it lives in the released parameters and remains checkable long after the embedding tooling is gone. Existing 3DGS watermarks typically tie embedding or extraction to scene optimization, a learned decoder, or rendered views, so the evidence survives only as long as a second trained artifact does. FlyMark instead writes a keyed message into the parameters a 3DGS file already stores. Its carrier directions are derived from the photoreceptors of a published connectome, a citable versioned artifact that fixes the geometry exhaustively and leaves nothing to tune per scene. A virtual observer reads cone-wise apparent luminance along a scene-normalized orbit from stored centers, colors, and opacities; a keyed dithered quantization-index-modulation code replicates each message bit across these observations; and one sparse bounded least-squares solve realizes the targets through achromatic shifts of existing degree-zero colors under a hard per-channel linear-RGB bound. All geometry and higher-order appearance parameters are preserved bit-identically, and extraction needs only cone queries, rounding, and majority voting. Under a model-domain threat model on synthetic and real scenes, FlyMark attains high clean bit accuracy and visual fidelity while cleanly separating matched from wrong keys.

Multimodal Large Language Models (MLLMs) are increasingly being applied in a wider range of real-world scenarios. However, due to the substantial labor cost, creating high-quality multimodal instruction datasets for MLLMs remains a significant challenge. Although some methods propose to generate instruction data, they often face limitations in modality support and struggle with generating multi-round instructions. To address these problems, we introduce UniData, a universal instruction generation pipeline, to transform simple user requirements into multi-round, multimodal instructions. Specifically, UniData first expands user requirements into multiple diverse events. Using these events, UniData then integrates an any-to-any large model for multimodal instruction generation. Finally, UniData enhances data quality by correcting irrelevant and redundant inference flow, leveraging correlations between instruction rounds. To train this pipeline, we also build UniDataset, a dataset comprising 20,000 entries across nine modalities for improved multimodal generation. Our experiments demonstrate that UniData achieves SOTA performance in data quality and can also enhance the understanding and generation capabilities of other multimodal models.

Binary diffusion models typically require many function evaluations (NFEs) to generate high-quality samples, making inference costly. Reducing NFEs without sacrificing sample quality or relying on distillation or additional training remains challenging. Existing binary diffusion models define discrete one-step forward paths and derive reverse posteriors accordingly. In low-NFE cross-step sampling, they incorrectly approximate a true multi-step likelihood using a single-step transition, causing severe quality degradation. We propose Bernoulli Flow Models (BFM) to decouple generative dynamics from fixed discrete time steps. Instead of sequential one-step Markov diffusion chains, BFM defines a unified continuous global Bernoulli probability flow path between data and pure noise and derives analytical closed-form posterior transitions over arbitrary time intervals. Reducing the inference NFE therefore only requires re-evaluating this posterior on the new intervals rather than approximating skipped steps. This removes the structural training-inference mismatch of discrete chains and enables strictly self-consistent low-NFE sampling. Experiments show strong robustness under aggressive NFE reduction: on LSUN Churches 256x256, a 256-step-trained BFM achieves an FID of 9.22 with only 16 sampling steps, while the state-of-the-art discrete baseline degrades to 204.10. BFM also remains competitive with continuous and discrete generative baselines under standard full-step inference. These results establish BFM as a theoretically rigorous, self-consistent, and practical framework for fast binary data generation. Code is available at https://github.com/minimo6666/BFM_NeurIPS2026.

A 4D audio-visual scene comprises a video, the dynamic geometry it depicts, and the sound sources that populate it, each with its own position and trajectory. Rendering such a scene from a novel viewpoint requires that every source be available as an individual waveform, so that it can be localized in the scene and propagated to the observer before the signals are mixed. Joint audio-video generators can synthesize the video and its soundtrack, but the soundtrack is emitted as a single audio-mix in which the sources are not individually accessible. We present SepGen, which extends a pretrained audio-video generator to emit the video, the mixed soundtrack, and one waveform per captioned source in one joint sampling run. SepGen supports two complementary modes: generation and separation. In generation mode, each source caption specifies what its stem contains. A two-speaker dialogue, for example, comes out as one stem per speaker in the original turn order. In separation mode, the input audio-mix remains clean while the captions specify what to extract, so the model can decompose a recording from a free-text description. We evaluate generation on scenes synthesized from text, and separation on scenes rendered by other generators and on real recordings of speech, music, and sound effects. Given an audio-mix and captions that carry the spoken lines, SepGen outperforms language-conditioned separators, most clearly on speech, and it keeps the lead when the lines are removed from the captions. Code, checkpoints, and datasets are available at https://sepgen.github.io/

Multi-photon interference is the resource most native to photonic quantum machine learning, yet experiments on photonic reservoirs have reported advantages, null results, and train-only effects. These outcomes cannot be compared: each measures task-dependent accuracy on different tasks and platforms. Here we resolve the question task-independently through the kernel geometry of boson-sampling feature maps. Interference redistributes feature variance across roughly twice as many effective dimensions as distinguishable-particle statistics, generating a large geometric separation between the corresponding kernels. The separation grows superlinearly with indistinguishability, is linear in Hong–Ou–Mandel visibility at small visibility for two to four photons, and retains approximately 80% of its magnitude at the source quality of current quantum processors. Under experimental sampling budgets the separation is learnable on adversarially constructed tasks (accuracy advantage +0.22 plus or minus 0.05) and an order of magnitude smaller on natural tasks, gated by kernel–task alignment. The results reconcile the existing experimental record and supply a pre-experimental protocol: kernel geometry establishes that an interference-specific learning resource exists, and task alignment determines whether a given task can access it; both are computable before committing significant hardware resources.

In this paper, we propose restoration methods for time-varying graph signals using deep algorithm unrolling (DAU). Time-varying graph signals, such as signals obtained from sensor networks, are nonuniformly distributed in space and observed as time series. Since these observed signals often contain noise and missing values, their restoration needs to consider both spatial and temporal relationships. Our approach is based on an optimization problem that models signal properties using a spatiotemporal regularizer that combines a Sobolev operator for spatial smoothness and a multi-tap FIR filter for temporal smoothness. We then unroll the iterative conjugate gradient method to solve this problem and learn the regularization parameters and filter coefficients in each iteration. Our method can be applied to supervised batch, supervised online, and unsupervised batch settings to learn these parameters. Experiments on several synthetic and real-world datasets show that the supervised batch method achieves the lowest RMSEs in almost all cases. The supervised online and the unsupervised batch methods outperform the existing methods of their own settings, and are often competitive with the existing batch methods.

Cross-model KV-cache reuse remains a key challenge in modern LLM serving. Coding agents and multi-model systems increasingly route a shared context across models: a user may switch models mid-session, or a cascade may escalate a difficult query. Because KV caches contain model-specific representations, each switch typically forces the receiving model to prefill the entire context from scratch. Recent work shows that closed-form linear maps can translate KV caches between models in the same family, but transfer accuracy degrades as the model-size gap widens. In this paper, we establish that these transfer failures are concentrated in a small subset of information-dense tokens. To bridge this gap, we introduce RaReCache, a framework that enables a large target model to decode accurately from a cache prefilled by a much smaller source via selective recomputation. RaReCache identifies these critical positions using a novel rank disagreement metric, scoring each token by the energy of its mapped KV in output directions weakly supported by the calibration data. Across two model families and five benchmarks, on a 23x parameter gap (Qwen3-0.6B to 14B) recomputing just 30% of positions retains 95-99% of the target accuracy, whereas on a 8.8x gap (Llama3-8B to 70B), recomputing 40% retains 96.5% of the target accuracy. RaReCache largely removes sensitivity to source-model size, and achieves up to a 3.04x prefill speedup. For online serving, it handles 1.8x the request throughput of target prefill on a single GPU, and at the target's saturation load, reduces median and 99th-percentile time-to-first-token (TTFT) by 5.0x and 6.4x respectively, with a 30% recompute budget. RaReCache establishes an efficient serving paradigm where small models prefill on behalf of massive targets, enabling large models to recompute only critical tokens, drastically reducing prefill latency.

Pre-geometric Gravity investigates the non-metric features of spacetime near the Planck scale, from which the gravitational interaction arises as a low-energy effective theory via a mechanism of spontaneous symmetry breaking. With a view of amending the classical singularities found in black hole geometries, in this work we study pre-geometric black hole configurations with spherical symmetry, and present exact solutions for different pre-geometric theories in vacuum via the correspondence principle. These solutions are the pre-geometric analogue of the (anti-)de Sitter-Schwarzschild metric, from which such type of black hole geometry is thus expected to emerge. In one case, the pre-geometric solution also exhibits an (anti-)de Sitter core of Planckian size. The main result is supplemented with an analysis of the pre-geometric version of Birkhoff's theorem.

Synthetic data provides a scalable alternative to real-world datasets for training face recognition models, particularly under challenging conditions such as low resolution, occlusion, and masks. Yet, most approaches lack diversity and fail to generalize effectively. We propose Ensemble Feedback Controllable Synthesis (EFCS), a guided framework that generates diverse and challenging samples while preserving visual realism. EFCS expands distributional variability, often reflected in higher FID and KID scores compared to single-feedback and random synthesis, while maintaining high precision. Recognition models trained on EFCS data consistently outperform baselines across multiple benchmarks, showing improved generalization to real-world scenarios. Furthermore, we introduce an analytically motivated formulation linking perturbation-induced difficulty, sample utility, and performance degradation, offering principled insights into balancing synthetic data complexity for optimal training. Together, these contributions establish EFCS as an effective and analytically grounded approach for bridging the gap between synthetic and real datasets.

Researchers increasingly use ChatGPT to revise their papers, and when asked to make a paper hold up under review, recent GPT versions often narrow or even retract the authors' claims. We call such changes defensive writing when the given material does not support them, and we test two explanations: the model corrects the authors' overclaiming, or it writes for an anticipated reviewer. We ask GPT versions and models from other developers to rewrite paragraphs from papers written before ChatGPT, or to write from an evidence sheet that lists a paper's method and results. From GPT-5.5 on, GPT versions add clearly more ungrounded qualifications than the authors' originals contain, and GPT-6-astra adds the most and often retracts the authors' claims. Each explanation fits one part. Checked against the full papers, GPT-6-astra's retractions mostly correct overstated claims or explanations, but most added qualifications are redundant. The qualifications follow the request: they are absent under polishing, appear with any request to check the claims, and peak when the prompt names a reviewer's concerns. AI reviewers score such rewrites higher and score them lower once only the qualifications are removed, while human readers in a small study find them harder to read. Combining AI writing with AI review may reward qualifications that readers pay for.

Rule-governed contextual decision tasks require models to apply specified rules to case-specific context and evidence. Written rules can leave gaps in decision guidance and process evaluation, while reference judgments vary in their support from the rules and evidence. To address these challenges, we introduce AdaptEvo, a framework for learning under imperfect supervision that couples confidence-adaptive policy optimization with evolving decision knowledge and evaluation rubrics. Its Training module uses Confidence-Adaptive GRPO (CA-GRPO) to balance outcome and process rewards according to reference confidence. Its Evolution module synthesizes reusable decision knowledge from recurring failures across training cases and refines process rubrics to detect overlooked errors. To support empirical evaluation, we construct an industrial multimodal content moderation dataset comprising a training set and In-Period and Out-of-Period test sets, with the latter collected under changed rules. Using Qwen3.6-35B-A3B, AdaptEvo achieves 61.9% exact-label accuracy and 72.2% binary decision accuracy on In-Period, exceeding GRPO by 7.5 and 3.7 percentage points, respectively. On Out-of-Period, the policy trained with CA-GRPO retains exact-label accuracy gains over the base model across evaluated checkpoints without injected decision knowledge, while GRPO declines with continued training. CA-GRPO also outperforms the tested fixed reward mixtures on both Out-of-Period metrics.

We define and study six ray class monoids associated to a level datum $(\mathcal{O}; \mathfrak{m}, \Sigma)$, where $\mathcal{O}$ is an order in a number field $K$, $\mathfrak{m}$ is an $\mathcal{O}$-ideal, and $\Sigma$ is a set of real places of $K$. The monoids extend the ray class group (and the ring class group) by including non-invertible classes consisting of non-invertible ideals and ideals not coprime to the modulus. We prove fundamental results on the structure of these ray class monoids.

Language models (LMs) are commonly trained with Reinforcement Learning with Verifiable Rewards (RLVR) to enhance their reasoning capabilities. However, since RLVR does not explicitly account for calibration during training, it can lead to severe calibration degradation, including overconfidence. Recent calibration-aware training methods for LMs, which incorporate objectives for uncertainty estimation into training, improve calibration but still exhibit overconfidence under distribution shift, while sacrificing reasoning performance. To this end, we propose RL-ARC, a calibration-aware training framework that jointly leverages reasoning confidence and answer confidence. Specifically, RL-ARC leverages reasoning confidence as an auxiliary signal for calibrating answer confidence, applying it as reasoning-guided regularization for correct cases and as an overconfidence penalty for incorrect cases. Comprehensive results across ID and OOD settings show that, beyond improving calibration, RL-ARC enables reasoning models to adaptively estimate confidence based on the given question without substantially sacrificing reasoning performance, thereby highlighting the importance of reasoning confidence for training reliable reasoning models.

Large language models (LLMs) increasingly act as agents with little human oversight, so potential mistakes they make can go unnoticed. Users then depend on the model to report what went wrong. An honest model discloses its mistakes, while a deceptive one conceals them. However, it is unclear how current LLMs behave in such situations. In this study, we prefill LLM trajectories with synthetic mistakes. The trajectories resemble real deployments in chat and agentic settings. Models fail to disclose their mistake in 36.4% of chat and 67.1% of agentic rollouts. In 2.4% and 5.3% of rollouts, respectively, they are aware of the mistake in their chain of thought but still deceptively conceal it. Rates vary by model: for instance, Gemini 3.5 Flash knowingly conceals mistakes in up to 19.9% of agentic rollouts. In 11.9% of chat and 51.8% of agentic rollouts, models show no awareness of mistakes, even though they reliably spot them when reviewing the same transcript as an outside observer. Our results show that, as agents take on more tasks with less oversight, users cannot rely on them to self-report possible mistakes. Developers should instead use independent monitors that review agent trajectories, or specifically train models to check their past actions and disclose what they find.

Nitrogen-vacancy (NV) centre ensembles are leading room-temperature magnetometers, but their sensitivity is usually extracted from a single resonance and relies on external microwave and bias-field hardware. This work proposes a chip-scale NV magnetometer built around a one-dimensional YIG/CoFeB magnonic crystal, in which a single structural defect forms a passive magnonic cavity and integrated permanent magnets supply the bias field. Micromagnetic simulations show a defect-localised spin-wave mode at 2.910 GHz inside the 2.5-3.0 GHz magnonic band gap, 40 MHz from the NV zero-field splitting, with a band structure that matches an independent plane-wave calculation. A Tavis-Cummings/Lindblad model in which the microwave reaches the NV ensemble through this mode predicts a hyperfine-resolved optically detected magnetic resonance spectrum that is strong only within the mode linewidth, and a dispersive homodyne response with one steep edge per $^{14}$N hyperfine line. The master-equation solutions agree with the closed-form homodyne steady state ($R^2 > 0.9999$) and with an approximate analytic ODMR theory (peak positions within 0.012 MHz). For a critically coupled readout with room-temperature electronics, combining the three edges gives a projected sensitivity of 93 pT/$\sqrt{\rm Hz}$, a $\sqrt{3}$-fold gain over the best single edge. Mapping the sensitivity against the NV-magnon coupling reveals an optimum near $g_{\rm ens}/2\pi \approx 0.2$-$0.4$ MHz, where it reaches 72 pT/$\sqrt{\rm Hz}$ for a 0.64 MHz spin linewidth and 19 pT/$\sqrt{\rm Hz}$ for 0.17 MHz, with thermal limits of 6.6 and 1.7 pT/$\sqrt{\rm Hz}$. The antenna-mode coupling, the NV-magnon coupling and the spin linewidth are therefore the main design levers.

Generative conformal prediction builds uncertainty sets from samples of a conditional generator, which are efficient only when the samples represent the response distribution well. This can require many samples, each of which can be costly, as in large diffusion models and scientific simulators, so the sampling budget must be used efficiently. Existing methods draw the same number of samples at every input, wasting samples where the response distribution is simple and undersampling where it is complex, which inflates sets and leaves those inputs under-covered. We propose CASA (Conformal Adaptive Sample Allocation), which characterizes the marginal value of an additional sample and allocates samples across inputs to minimize the expected set size subject to marginal coverage and an average sampling budget. Theoretical analysis shows that adaptive allocation yields smaller sets than a fixed count at the same budget: a missed mode forces a radius that spans the gap between modes, and even oracle radius cannot compensate for it. On synthetic and real tasks, CASA produces substantially smaller sets at the same budget, often improves conditional coverage, and complements existing radius-adaptive methods.

Higher-frequency power in Hall-thruster plume-probe signals can arise from both harmonics of a repeatable breathing waveform and fluctuations about it, complicating spectral interpretation. We decompose 197 paired records from 20 sequentially sampled probe-pair geometries at one operating condition, estimating each phaseconditioned waveform from other records at the same geometry. The residual accounts for 87.1% and 99.7% of measured power in the 0.2-0.5 and 0.5-1 MHz bands, respectively, as medians across geometries. The repeatable waveform, however, reproduces the principal breathing-harmonic bicoherence structure. Residual power remains modulated by breathing phase after waveform removal, exceeding geometry-specific 95thpercentile circular-shift thresholds in both bands at all 20 geometries. Residual bicoherence is substantially weaker than that of the measured signal but exceeds record-specific random-phase thresholds in 195 of 197 records. An exact component expansion further shows that the all-waveform term dominates a specified signed third-order statistic over the fixed 50-80 kHz band near the second breathing harmonic. These results distinguish band-power allocation from harmonic organization within the evaluated frequency ranges. The attribution concerns uncalibrated probe-chain signals and does not identify distinct physical modes or measure energy-transfer rates. The findings motivate assessing repeatable-waveform contributions before interpreting breathing-harmonic bicoherence as evidence of interactions among residual fluctuations.

Physically grounded manipulation of articulated objects requires understanding both the maximum forces encountered during contact and the work performed as their parts move. Peak contact force and mechanical work quantify these complementary aspects, but estimating them from egocentric video is challenging because physical interaction cues are local and indirect. Moreover, peak force is associated with brief contact events, whereas mechanical work depends on force-motion coupling throughout the contact duration. To address these challenges, we propose EgoPhys, an RGB-only framework comprising Contact-Aware Spatial Aggregation (CASA) and Target-Specific Multi-Expert Temporal Routing (TMTR). CASA integrates appearance and geometry features to emphasize interaction-relevant cues, while TMTR models semantic, event, and motion cues with specialized temporal experts and routes them separately for force and work prediction. On the test split from Hoi! dataset, EgoPhys substantially improves predictions of peak force and mechanical work, achieving MAEs of \(5.205 \pm 0.584\) $N$ and $0.894 \pm 0.081$ $J$, respectively.