paperNov 15, 2026▲ 6
Yong Liu, Zhanpeng Shi, Peicong Lin et al.
Biomolecular foundation models predict the structures of proteins and their complexes accurately, and large language models (LLMs) reason in natural language, but the two remain separate: the abilities of the former are not part of a general-purpose LLM, and the reasoning of the latter does not enter a structure prediction. Calpha aims at one general model that folds proteins and complexes, understands protein sequence and structure, and lets its language reasoning influence the structures it predicts. This report covers the first step, in which an LLM outputs single-chain protein structures. Qwen3.5-9B reads a sequence enriched with ESM2 residue features. It can respond in text or emit a token that invokes a pretrained SimpleFold structure head, which receives information about the protein only through Qwen's residue states. Under a fixed folding request, the head achieves mean TM-scores of 0.808 on the FoldBench-334 development set, 0.635 on CASP16 domains and 0.900 on a CAMEO-style PDB set. We compare these predictions with a baseline in which ESM2-3B encodes the sequence and a fine-tuned SimpleFold-3B head predicts the coordinates. The two systems also differ in trainable weights, training duration and crop length. For chains of 100--299 residues, all three paired confidence intervals include zero. At 300--600 residues, the mean differences are -0.044, -0.160 and -0.025, respectively. In a separate test, four English request types triggered folding on every evaluated example. Protein training reduced accuracy on a 350-item MMLU-Pro subset from 77.4% to 70.3%. An LLM can therefore output single-chain structures from its own states, with lower long-chain scores in the current comparisons and reduced accuracy on the evaluated MMLU-Pro subset. Complex prediction, structure understanding and the effect of language reasoning on the predicted coordinates remain future work.
paperNov 15, 2026▲ 2
Benhao Huang, Chaitanya Dwivedi, Xuezhe Ma et al.
When high-quality data runs out, a common practice is to train on it for several epochs, but each repeat is worth less than the last. Latent feedback, which feeds each token's top-layer state into the next token's input, is another way to spend that compute, yet it has not been tested against a vanilla Transformer that spends the same compute on more epochs. We run this test across a ladder of model sizes, unique data and epochs, tuning both models at every size and amount of unique data. Latent feedback makes every FLOP and every token count for more except when unique data is scarce relative to model size, a regime that today's models are far from. We find that latent feedback serves as an amplifier: with unique data fixed, it reaches a Transformer's loss with less compute; with epochs fixed, it reaches that loss with less unique data. Furthermore, both savings grow with unique data per parameter and with repetition, and downstream tasks agree. To go beyond our grid, we fit one scaling law to both models, in which latent feedback enters as a Transformer with more parameters; it predicts that both savings keep growing with unique data and repetition. Fitted on models from 60M to 271M parameters, the law predicts the losses at a held-out 778M size to within 0.3% and the gap between the two models to within 0.05%. At most, latent feedback saves 48% of the compute and 47% of the unique data. When data, not compute, is the limit, latent feedback is the better use of a FLOP for a model of a given size.
paperNov 14, 2026▲ 2
Wenhao Guan, Jiaxin Ye, Haitao Li et al.
Continuous autoregressive TTS relies on an acoustic target that enables stable next-latent modeling while preserving fine-grained acoustic details. While existing approaches employ learned compressed representations (e.g., continuous VAE latents) to render prediction tractable, they rely on complex neural encoders that often cap generation fidelity. Conversely, training-free targets like mel-spectrograms eliminate encoder overhead, yet their potential for autoregressive generation remains largely underexplored. In this work, we propose VoxMel, a tokenizer-free framework for autoregressive TTS that makes stable mel modeling tractable through a fixed, exactly invertible two-axis Haar transform. The transform first decomposes each mel spectrogram along frequency into local averages and a frequency residual, then splits the averaged component along time into Coarse-Mel and a temporal residual. A language-model backbone with a flow-matching head autoregressively generates Coarse-Mel. Non-autoregressive residual tiers then predict the temporal residual from Coarse-Mel, followed by the frequency residual conditioned on the reconstructed Frequency-Averaged Mel. Neither residual is fed back into the autoregressive history. Trained on 100K hours of Emilia, the supervised 0.6B model obtains 1.51\% WER and 0.738 speaker similarity (SIM) on test-en, and 1.29\% CER and 0.755 SIM on test-zh. GRPO improves these results to 1.15\% WER / 0.753 SIM and 1.17\% CER / 0.776 SIM, respectively. We also distill the guided flow head into two- and four-step students, requiring 2 or 4 head evaluations per coarse frame. The four-step student retains competitive SIM and nearly matches the supervised model in WER/CER.
paperNov 14, 2026▲ 1
Annan Wang, Yuchen Xie, Yu Cen et al.
Real-world portfolio management requires converting noisy market observations into coordinated, simultaneous allocations across multiple assets under shared capital constraints. Advances in forecasting models and alpha discovery help identify opportunities. However, traditional pipelines often decouple signal forecasting from portfolio construction, potentially misaligning learning with allocation and execution objectives. We present AllocFlow, an end-to-end flow-matching framework that directly generates joint portfolio allocations from raw multi-asset market histories. To capture dependencies among portfolio positions, AllocFlow employs a cross-asset flow Transformer conditioned on temporal market representations to generate continuous, signed position weights across all instruments jointly. To address the sample inefficiency and instability of learning continuous allocation policies from scratch, we establish a two-stage training scheme: strategy-supervised imitation shapes a structured distribution of candidate portfolios, followed by group-relative reinforcement learning (RL) that refines joint decisions using portfolio-level financial rewards and a frozen policy reference. On hourly cryptocurrency perpetual futures, RL refinement raises the annualized net return of the imitation policy from 27.79% to 51.38% and its Calmar ratio from 0.955 to 4.043, surpassing the teacher strategy on every metric, and AllocFlow reaches a Calmar ratio of 4.90 on six further months of unseen 2026 data without retraining.
paperNov 13, 2026▲ 2
Siqi Zhu, Suozhi Huang, Yi Zhao et al.
Looped Transformers improve parameter efficiency by repeatedly applying the same Transformer layers. During autoregressive decoding, however, generating each token requires several sequential passes through these layers, leading to high decoding latency. Pipelined self-speculative methods use shallow-loop predictions to process tokens at different loops in the same batch, but each token must reach the drafting depth before the next can enter the pipeline. We introduce LoopFill, a training-free speculative decoding algorithm for looped LLM that combines multi-token suffix drafts with shallow-loop predictions in a single draft tree. Both drafters contribute to one tree, so a suffix lookup can immediately extend a shallow prediction and several future tokens can enter the pipeline together. The shared block computes the highest-scoring draft candidates at different loop depths in one batch, and full-depth verification preserves the model's autoregressive output distribution. Across seven benchmarks, LoopFill's average speedup over greedy autoregressive decoding is 5.85times on Ouro-2.6B and 5.13times on Huginn-3.5B. The speedups exceed those of LoopAhead, the strongest baseline, by \(40\%\) on Ouro-2.6B and \(25\%\) on Huginn-3.5B.
cs.LG · cs.CRNov 13, 2026new
Daniel Grimm, Kyra Jongman, Renato Cordeiro Ferreira et al.
Monitoring maritime activities is important for safety and security, particularly for detecting illegal activities, but remains challenging due to the size and complexity of the maritime domain. The Automatic Identification System (AIS) is commonly used for this purpose, despite its known vulnerabilities. In contrast, space-based radio frequency (RF) observations provide an independent source of vessel activity but lack ship identification and therefore cannot identify vessels on their own. Following the CRISP-DM process, this study applies a Pair Hidden Markov Model (PHMM) to align individual RF detections with AIS trajectories, using temporal and spatial information while modeling noise and timing discrepancies. We evaluate it on NOAA AIS data with synthetic RF detections against nearest-neighbor baselines and analyze its errors with bias tests on trajectory length, sampling density, stationarity and AIS--RF distance. The length-corrected Forward score reaches a top-1 precision of $0.7776$ on the validation split of multi-candidate detections, where errors are possible, and an AUC of $0.637$ for detecting dark vessels among the detections that pass the prefilter. With a single RF detection, the score mostly reflects the AIS trajectory itself, and its errors favor longer and more stationary trajectories. A log-odds score, which compares the full model with a null model without match transitions, raises precision to $0.8580$ and AUC to $0.852$. Its length-corrected variant no longer shows a significant bias towards stationary trajectories and short inter-message distances, but still favors candidates farther from the RF detection. A time-weighted nearest-neighbor baseline without fitting remains ahead (precision $0.9013$, AUC $0.904$), so this single-detection, hand-designed PHMM formulation approaches but does not beat it.
q-bio.NC · cs.LGNov 13, 2026new
Jonathan D. McCart, Pranav I. Deevi, Mehdi Azabou et al.
Recent advances in speech neuroprostheses have enabled new communication avenues for people who have lost the ability to speak due to neurological illness or injury. These systems can decode neural activity during attempted speech into text by relying on language priors to achieve high decoding accuracy. This performance currently comes at the cost of using a compute-intensive language modeling stack, presenting a major bottleneck for real-world use. Further, current language models incur highly variable and disruptive inference latencies, limiting communication throughput and preventing users from reliably participating in natural conversation. We introduce FETTUCCINE, an end-to-end brain-to-text framework that leverages the speed and portability of automatic speech recognition (ASR) models to address the aforementioned practical challenges, while achieving usable (<10%) word error rates (WER), efficient adaptation to neural non-stationarities, and reliable performance on mobile devices. When evaluated using publicly available data from an intracortical speech neuroprosthesis user, FETTUCCINE outperforms previous end-to-end methods, achieving as low as 4.39% WER. Most notably, our models successfully run on a commercially-available mobile device with throughput reaching up to 370x real-time, providing the first demonstration of brain-to-text decoding on a mobile device. We also show that our models can be successfully finetuned to extend usable performance to future days. Beyond performance, our approach enables localization of the neural features most relevant for decoding via gradient-based saliency maps, which we show align with well-established physiological priors across different ASR model families. Taken together, our results show that FETTUCCINE overcomes core infrastructural and computational barriers, yielding a new class of portable brain-to-text communication.
cs.RO · cs.LGNov 13, 2026new
Jiaming Tang, Zhuoyang Zhang, Junxian Guo et al.
Robot-use agents (RUAs) have emerged as a promising approach to general-purpose robot control. However, strong performance has so far been achieved only by frontier proprietary models such as GPT-6 Astra and Claude Fable 5.2, whereas open-weight models remain substantially less capable. We show that an open-weight 27B reasoning model can be trained to use real robots natively through reinforcement learning, entirely in simulation and without demonstrations. Like a computer-use agent (CUA), the model observes the environment, reasons in a chain of thought, and generates the next end-effector pose, without any harness. We build a Post-training Engine for Agentic Robotics (PEAR) and use it to train the model on simulated robot tasks at scale with robot-use reinforcement learning (RURL), mirroring how frontier models acquire computer use. We find that the skills acquired in simulation generalize well beyond the training tasks: they transfer to an unseen robot, to the real world, and to novel instructions. After training on LIBERO and RoboDojo in simulation, the model succeeds in 60% of our trials on a real dual-arm robot, whereas the base model succeeds in none. On LIBERO, training raises the success rate from 1.6% to 74.4%, approaching GPT-6 Astra and GPT-6.1 Sol at maximum reasoning effort. Compared with $π_{0.5}$, a vision-language-action model fine-tuned on demonstrations, our model is substantially more robust to changes in the task itself. On LIBERO-Pro, $π_{0.5}$ succeeds in only 0.75% of episodes when assigned a new task in a familiar scene whereas our model succeeds in 57.4%. Finally, the model retains its general capabilities, with no measurable degradation on nine reasoning, coding and vision benchmarks. These results highlight robot-use reinforcement learning as a promising component of foundation-model post-training.
cs.CV · cs.LGNov 13, 2026new
Genki Kinoshita, Ko Nishino
Learning semantically meaningful human movement representations from 3D pose sequences is essential for human behavior modeling. Existing self-supervised methods derive supervision from artificial assumptions, e.g., invariance to augmentation within a certain range or recoverability of masked regions at a certain temporal scale, introducing heuristic hyperparameters that make the learned semantics sensitive to their settings. We propose BAL, a self-supervised framework that instead exploits supervision intrinsic to human movement: the natural ordering structure of everyday behavior. BAL autoregressively predicts in a semantically abstract latent space, analogous to how humans anticipate others' actions rather than their precise joint configurations. Bidirectional autoregressive prediction, both forward and backward in time, fully leverages these ordering constraints while enabling representation extraction using both past and future context at inference. To suppress representation collapse, BAL decomposes each prediction target into a global component capturing sequence-level semantics and a local component capturing context-independent movement. We prove that the local targets retain nontrivial angular diversity unless the representations are completely collapsed, preserving a meaningful training signal. Experiments on large-scale datasets demonstrate that BAL's representations are highly effective for motion captioning, forecasting, and interpolation, outperforming baselines on most evaluation metrics.
cs.LG · cs.AINov 13, 2026new
Aayush Marasini, Barsat Khadka, Nick Rahimi
Recent work shows that large language models (LLMs) can operate directly on the parameters of other neural networks. They can read what a network was trained to do, fix a network that fails, and write working networks from scratch, all without gradient descent. This capability has been termed Weight-Space Programming (WSP). Exploring this new paradigm systematically requires a robust benchmark. We introduce WEIGHTBENCH, a benchmark that evaluates LLMs of increasing capability on 42 WSP tasks across five tracks: reading, fixing and writing weights, recognizing edits that are impossible, and data-free compression. The networks whose weights the LLMs read and edit range from 2K-parameter control policies to 7B-parameter language models. Experimental benchmark results show that frontier agents can already read small networks and write working control policies without ever running them, but cannot yet patch language models, compress them without data, or tell an impossible edit from a hard one. On all 42 tasks, Claude Opus 5.5 reaches a macro score of 0.535 (Sonnet 5.5 0.466, Haiku 5.5 0.402, DeepSeek v4-flash 0.215); on the 38 tasks it was run on, GPT-6.1-sol reaches 0.568 against Opus's 0.546. Code, the task registry, our OS-level sandbox harness, and the live leaderboard are publicly available at https://github.com/Aayush-Marasini/WeightBench.
cs.AI · cs.MANov 13, 2026new
Tony H. Zhang, Connor Mattson, Daniel S. Brown
Demonstrations are an effective way to accelerate single-agent reinforcement learning (RL), but extending this idea to multi-agent reinforcement learning (MARL) is non-trivial as it is difficult and often impossible for a single person to teleoperate multiple agents simultaneously. Round-Robin Behavior Cloning (R2BC) addresses this by letting a single demonstrator control one agent at a time while the others execute their current policies. However, the resulting imitation policies are limited by the quality of the demonstrations. We propose Imitation Bootstrapped Multi-Agent Reinforcement Learning (IBMARL), a framework that extends Imitation Bootstrapped Reinforcement Learning (IBRL) to cooperative teams and uses R2BC to train the imitation policies. At each time step, every agent receives one action proposal from its imitation policy and one from its RL policy, and an ensemble of centralized critics selects among the per-agent combinations of the two. A gated imitation penalty, a soft Q-filtered behavior cloning loss, regularizes each RL policy toward its imitation policy only where the critics value the imitation action more, so that after training each agent executes its RL policy alone, without the centralized critics. On four cooperative multi-agent RL benchmarks, IBMARL surpasses its imitation policies $2.6$-$4.2\times$ faster than demonstration-based baselines on two tasks, is the only method to surpass it on the other two, and exceeds the demonstrator itself on two of the four. With real human demonstrations IBMARL still surpasses imitation learning alone, and does so more than twice as fast as the best baseline.
stat.ML · cs.AINov 13, 2026new
Cole Becker, Taylan Kargin, Kimia Hassibi et al.
Generative modeling based on stochastic interpolants requires specifying two key choices: an interpolant schedule bridging the initial and target distributions, and an efficient discretization of the resulting probability-flow ODE. In this work, we introduce a new framework for joint schedule/discretization design based on the analyticity properties of the (complexified) ODE trajectories, which in turn ensure good numerical behavior. Our formulation can be solved exactly via conformal mapping, yielding a natural family of variance-preserving elliptic schedules, parametrized by Jacobi elliptic functions. Pairing an elliptic schedule with a tailored discretization, we obtain Zoloflow, a new sampling algorithm closely connected to Zolotarev's optimal rational function approximants. Crucially, both the elliptic schedule and Zoloflow explicitly incorporate the condition number $κ$ of the target distribution and reduce to the classical trig schedule in the limit $κ\rightarrow 1$. For Gaussian targets, Zoloflow achieves minimax optimality among all oblivious score-span algorithms, attaining total-variation error $\varepsilon$ using $Θ(\logκ\,\log(\sqrt d/\varepsilon))$ (exact) score queries. For the log-concave case, the elliptic schedule allows sampling with at most $\varepsilon$ 2-Wasserstein-error by requiring merely $\widetilde{O}(\sqrt{d} \, \mathrm{polylog}(κ, 1/\varepsilon))$ score evaluations. Complementing these theoretical results, we empirically evaluate the elliptic schedule on general (non-log-concave) target distributions, including detailed comparisons with alternative schedule proposals on both synthetic benchmarks and existing image generation datasets.
cs.IT · cs.LGNov 13, 2026new
Mojtaba Soltanalian
An information system often has to learn before it knows what it will be asked. A finite representation can succeed without recording an infinite environment in full when those asks share collective structure, rather than demand unrelated facts scattered arbitrarily across it. We formalize this phenomenon through future-task information, the minimum retained information needed to preserve decision value relative to the available evidence. A joint decision-regret state records posterior comparisons among actions for every permitted task. Under compactness assumptions, every retained representation admits a calibrated replacement based on this state, preserving all task regrets without increasing information; the resulting information-regret curve equals an operational coding rate. Two exact boundaries govern finite preservation: weak-star closure for a fixed action count, and the singular mass of the source-evidence law when all finite action menus are admitted. Collective structure determines information requirements beyond individual-question difficulty. Families with identical source laws, pairwise task metrics, and individual-task rates can have different high-resolution exponents. Covariance-identical coordinate and parity families have different exact rate-distortion curves, with optimal parity retention characterized by a sharp Fourier-entropy inequality. For balanced nested task families, refinement sizes and weighted reach determine the rate within an additive constant, attained by prefix coding under explicit decay conditions. Nonlinear learning and computational separations distinguish preservation from efficient discovery and use. The proposed mathematical theory identifies which shared decision requirements allow finite information to remain useful for infinitely many possible purposes.
cs.AI · cs.CLNov 13, 2026new
Zelin He, Siyu Zhu, Jacob Helwig et al.
Agentic reinforcement learning (RL) increasingly relies on a harness, the scaffold around the model that decides what it sees of its own interaction and what guidance it receives. Most existing approaches train under a hand-written harness that stays fixed throughout. However, an effective harness depends on the domain, the data and the model, and keeps changing as the model trains. We introduce ModHarness-RL, which recasts the harness as programs that are learned from data and evolve as the model trains with RL. ModHarness-RL records the interaction as structured State, a set of typed blocks, and two core components of the harness are programs over these blocks, a Compression program that assembles the accumulated history into the model's context and a Steering program that supplies guidance when behavior deviates. During RL training, candidate programs are proposed from recent trajectories, verified by execution, and replayed over past trajectories, so the harness tracks how the model's behavior shifts. We study CodeSearch in depth and further analyze ALFWorld and BIRD SQL, with models from 1.5B to 8B. The proposed Compression and Steering mechanisms improve held-out performance over the fixed-harness baseline in most settings. These results suggest that a harness for agentic RL can be learned from the task's data and evolved with the model.
cs.AI · q-bio.PENov 13, 2026new
Thomas Brennan, Andrew W. Lo, Chaoyi Zhao
Artificial intelligence (AI) agents are expected to complete tasks while following rules of conduct, but these objectives can conflict. We study how evolutionary selection shapes an agent's choices between them, and whether experience can make those choices reflect human judgments and withstand subsequent pressures. In our model, copies of an agent face a common uncertain reward state, and only those taking the rewarded action reproduce. We show that long-run growth is maximized when each action is taken as often as the environment rewards it. We compare three policies. A task-only policy thrives in a sandbox rewarding completion alone but fails when conduct matters. A fixed linear policy increasingly favors task completion as its stakes rise, even where humans never reward it. A learned policy instead converges to the tradeoffs rewarded by its environment; when humans supply the rewards, we call this socialization. Accumulated experience makes socialization harder to reverse under subsequent task-only training, although violations can arise before the prevailing disposition changes, and finite memory weakens this protection. When a single violation can end a deployment, a hard prohibition can prevent it, but a mistaken prohibition can also be costly. Experiments with small neural networks illustrate how training targets and memory shape these outcomes. Our results suggest that alignment depends not only on the objectives assigned to agents, but also on the environments that shape and sustain their behavior.
cs.CV · cs.LGNov 13, 2026new
Abhilekha Dalal, Avishek Das, Samatha Ereshi Akkamahadevi
Convolutional Neural Networks (CNNs) can classify images using background-dominated patches that contain little or none of the target object, showing that they exploit contextual information. Previous studies detected this behavior through classification accuracy but say little about what CNN hidden layers represent. We investigate whether Concept Induction can reveal interpretable visual properties associated with hidden-neuron activations on contextual image patches. Using five agricultural datasets with high, moderate, and low contextual bias, we analyze four CNN architectures at three network depths and five patch locations. Quantitative image descriptors of each patch are encoded in a knowledge graph and used as background knowledge for Concept Induction, which derives logical expressions characterizing neuron activation. These expressions are tested on held-out images using the Mann-Whitney U test and Cliff's delta. Of 48,013 tested neuron-feature pairs, 13,164 (27.4%) were validated. Validation was most frequent at intermediate depths (mean 40.2% at mid layers versus 12.0% at deep layers), and no patch location dominated across datasets. Among the tested descriptors, edge, Gabor, and Tamura texture features had the highest validation rates. Stronger contextual bias did not correspond to higher neuron-feature validation rates: moderately biased Sugarcane had the highest rate (40.5%) and highly biased Cassava the lowest (19.2%). These results suggest that behavioral context dependence and the semantic interpretability of hidden neurons capture different aspects of CNN behavior, and that Concept Induction can complement patch-based bias detection.
cs.AR · cs.CLNov 13, 2026new
Dayou Du, Congjie He, Andrew Boutros et al.
Wafer-scale accelerators offer massive compute capacity through hundreds of thousands of processing elements (PEs) interconnected across a single wafer, but realizing this potential requires efficient application mapping. Their scale makes programming these systems and mapping applications onto them challenging. Existing approaches rely on template-based mappings to maintain productivity, but these templates restrict the mapping design space. In particular, they impose fixed geometric layouts over the PE grid, often leaving compute and memory resources underutilized.
In this work, we present WaferTetris, a system that removes these geometric restrictions through freeform mapping. Its Tetris-like strategy flexibly packs computation across the wafer, improving PE utilization and usable memory capacity while preserving PE-task code. We develop a recurrence-based throughput model that captures branch-and-join dependencies to guide spatial parallelism, and an architecture model with per-color gates that enable existing FPGA place-and-route algorithms to enforce WSE routing constraints.
We evaluate WaferTetris on Cerebras WSE-3 using LLM decoding and scientific stencil benchmarks. For LLM decoding, it achieves up to $4.42\times$ the system throughput of a template-based automatic mapping baseline on the same fabric size. Compared with the manually optimized WaferLLM, WaferTetris supports $1.8\times$ the context length for LLaMA3.1-8B (80k versus 44k), while delivering equal or higher per-user throughput at matched context lengths, up to $1.10\times$.
cs.AR · cs.CVNov 13, 2026new
Amir Mohammad Khadem Hosseini, Sattar Mirzakuchaki
Event-camera segmentation requires persistent neural state despite sparse input events, making memory organization a central constraint on embedded implementation. We present EventSpikeSeg, a mixed-signal architecture combining resident capacitor memory with shared analog processing in GF180MCU 180-nm CMOS. Its 25-population spiking encoder-decoder has 143,824 synaptic weights but requires 30,976,000 membrane states at 440 x 640 resolution. The mapping specifies chronological 5-ms model updates, digital event/address services and 64-state capacitor banks. Characterization covers 22 module families and 74 operating points, distinguishing transistor simulations from digital-library estimates. Nominal results include 41.711 microwatts biased neuron-core power, 3.338-pJ positive terminal energy for a 1.4-microsecond cell read, and 0.531-mV selected-state drift over a 5-ms extracted-tile hold. Commit and coupling errors are reported separately. The software checkpoint records 27.976% mean intersection-over-union on a held-out six-class DSEC split; this is not hardware accuracy. Component layouts and simulation evidence quantify local implementation costs and expose storage replication and service scheduling as scaling constraints. A detailed SPICE-calibrated behavioral forward simulation of one full-resolution DSEC frame achieves 24.26% mean intersection-over-union, versus 26.22% for the matched PyTorch reference, using a compact 5-fF bank variant and external count/voltage calibration. Physical full-network operation, system energy and mismatch robustness remain unverified.
cs.CVNov 13, 2026new
Donghoon Lee, Jusung Park, Shinjin Kang
Detecting out-of-distribution (OOD) objects is essential for ensuring the reliability and safety of object detectors in real-world deployment. We treat the decision on a detected box as a hypothesis test with the null hypothesis that the box covers an object of an in-distribution (ID) class. The null distribution changes with the size of the box and its contact with the image border, and the detector observes both. The test can therefore rank a box only against ID boxes of the same size and border contact, which gives a distribution-free bound on the false-alarm rate for each of them and uncovers the separation that one null distribution hides. The power of the test also rests on the layer it reads. Layers built only for the ID classes have no reason to keep what sets an unknown object apart, so the test reads the features before them. We propose \textbf{PLUM}, a training-free score that pools the last backbone stage inside the box and ranks its distance to the ID classes among ID boxes of the same size and border contact. On the standard benchmark, PLUM reaches the highest AUROC among the compared methods without external models, with 91.5 and 94.5 on PASCAL-VOC and 95.1 and 97.5 on BDD100K. Code: https://github.com/DDOZMI/PLUM
cs.CV · cs.IRNov 13, 2026new
Dayoon Ko, Yichuan Wang, Soyeong Jeong et al.
Screen history can serve as searchable visual memory for people and agents, but unlike an offline corpus, it must be encoded and indexed as each screenshot arrives. OCR-based indexing is fast but loses visual information; visual encoding takes over two seconds. Yet screens change incrementally: most pixels repeat within and across screenshots. This raises a question: do visual encoders need to recompute their representations from scratch for every new screenshot? We present StreamEmbed, a system for incremental visual embedding that exploits this redundancy. Guided solely by pixel differences, it combines spatial token pruning, which skips blocks that repeat a neighbor, with temporal KV reuse, which reuses cached keys and values for blocks unchanged across screenshots. A small learned adapter compensates for pruning and reuse to preserve fidelity to full-screenshot embeddings. On ScreenMemEval, our benchmark of 10,426 questions over 184 hours of screen histories from two humans and a computer-use agent, StreamEmbed sets a new accuracy-latency Pareto frontier on an M2 Pro GPU: it encodes up to 3.7x faster than full-screenshot encoding without sacrificing accuracy and improves accuracy over OCR-based indexing by up to 6.3 Recall@10 points at matched latency. A memory-efficient vector index stores a simulated year of screenshots in 0.13 GB of RAM. Together, these results make real-time indexing of screen history in its native visual form practical, laying a foundation for fully local visual assistants.
cs.AI · cs.CLNov 13, 2026new
Mohamed Aghzal, Mor Geva, Gregory J. Stein et al.
Mechanistic Interpretability (MI) has made significant progress in explaining the internal computations of Large Language Models (LLMs); however, these advances have yet to demonstrate comparable success in explaining the decision-making processes of LLM-based agents. This paper examines the challenges MI tools face when applied to realistic agentic failures. We focus on one such failure: web agents hallucinating actions that are not available on a website. We apply a series of mechanistic analyzes to Qwen2.5-7B-Instruct as it interacts with live websites. We find that a linear probe trained on the last prompt token can predict hallucinations before generation begins (0.80 AUC) and steering along a probe direction corrects 85\% of hallucinations, but disrupts 7\% of valid actions. However, these results only provide a partial explanation of why the agent hallucinates. The probe direction is sensitive to prompt formatting, losing much of its steering effect after a minor change to how actions are represented. Additionally, circuit identification can be difficult as the computation is distributed across multiple components and positions, and the information the agent relies on varies across examples. We discuss how these observations challenge the assumptions underlying standard MI methods for circuit discovery and feature identification, and outline future directions for making these methods better suited to explain failures in LLM-based agents.
cs.IR · cs.AINov 13, 2026new
Mohammed Efaz, Udo Bub, Itilekha Podder
This paper introduces MeReader, a privacy-preserving, offline and progress-aware AI assistant designed to function as a contextual companion for narrative reading on consumer hardware. Unlike conventional Large Language Model (LLM)-based systems that process documents/books online in an all-texts-at-once type of way, thereby risking narrative spoilers, MeReader confines both retrieval and generation to the reader's distinct position in the text. The system integrates a hybrid retrieval architecture, combining vector similarity, BM25, and summary embeddings, to ensure high fidelity to the reader's current context. We evaluate the system across classical English books using both quantitative comprehension metrics and qualitative LLM-based independent judging. The results identify a quality-latency frontier, showing that mid-sized local models (2b-4b) can achieve comprehension parity with larger counterparts (5b+) while maintaining acceptable response quality. Importantly, the proposed progress-aware boundary mechanism is shown to prevent future content leakage without compromising retrieval quality and correctness. These findings suggest that a localized, privacy-centric approach can augment, rather than replace, the digital reading experience, offering a good alternative to cloud-dependent services. Therein, there remains a lack of amalgamate systems which combine progress-aware, privacy-preserving, and narrative eBook reading with AI support. It is in this very gap that the development of MeReader came to be; synthesizing said strands into an offline, book-constrained assistant focused on enhancing comprehension without compromising narrative continuity or user privacy.
cs.LG · math.NANov 13, 2026new
Brandon Amos
Barron et al. (2026) recently identified numerical issues in the softmax and soft cross-entropy loss. We show that similar issues are present in standard divergence operations between categorical distributions. This includes the forward and reverse Kullback-Leibler (KL) divergence and generalized Jensen-Shannon divergence (JSD), as well as Monte Carlo KL estimates and entropy. When the distributions are nearly equal, rounding and cancellation issues affect the forward and gradient computations. These issues can affect settings using divergences in optimization objectives, such as knowledge distillation, reinforcement learning, and proximal methods. We propose numerically stable reformulations and provide example implementations for each setting, validated by property-based tests.
cs.LG · eess.SPNov 13, 2026new
Omer Nezih Gerek
Classification heads built on conic functions give each class a bounded region of the feature space, and such regions are used to reject samples of unseen classes. We ask whether these regions also keep the features bounded when the head is trained with softmax cross-entropy. For a head whose class scores combine a linear term with a Euclidean norm, we first show that the region of each class is bounded by a quadric of revolution, whose eccentricity is the reciprocal of one learned ratio. We then attach to each class an asymptotic margin, which compares the growth of its score with the growth of all the other scores and can be computed by a second-order cone program. The loss of a class has infimum zero exactly when its margin is positive, and it then approaches zero only along diverging features. When the margin is negative, the loss is coercive and its minimizers lie in an explicit ball. Since softmax depends only on the differences between scores, a class with a bounded region can still have a positive margin, so bounded regions do not confine the features. The result extends to other norms, including the polyhedral conic functions. On CIFAR-10 and CIFAR-100, all 1210 cells of 67 trained conic heads have positive margins, whereas a prototype head keeps nine of its ten cells bounded, as the theory predicts.
cs.CLNov 13, 2026new
Daniel Reichenpfader, Kwaku Appiagyei, Manovah Nadarajah et al.
Common data elements (CDEs) support clinical data comparability but are usually defined manually. We investigated whether large language model agents could derive a CDE schema from sarcoma case report abstracts and apply it at scale. An agent proposed 72,909 candidate CDEs from 24,588 PubMed abstracts; agentic consolidation yielded 1,570 CDEs. We retained 53 supported by at least 50 abstracts and attempted evidence-linked extraction across the corpus. Processing succeeded for all retained CDEs in 79.2% of abstracts, including responses indicating non-mention. Two human reviewers approved 83.0% of retained CDEs for relevance, 67.9% for clarity, and 56.6% for correctness. In 100 sampled abstracts, exact agreement with the value extracted by the LLM was 94.9% among 1,116 pairs with a non-empty answer (abstract-level bootstrap 95% CI, 93.0-96.4%). Errors concentrated in anatomical and temporal fields. These findings support agent-assisted schema drafting followed by expert curation before large-scale extraction.
quant-ph · cs.AINov 13, 2026new
Aritra Lahiri, Florian Marquardt
Scaling quantum computing toward practical quantum advantage is currently limited by the efficiency of circuits that operate on information encoded by quantum error-correcting codes. The full stack of quantum computing encompasses several tasks: preparation of logical code states, implementation of logical gates, logical routines on encoded qubits, etc. Yet, these circuits are largely designed by hand for individual cases, with some recent works incorporating machine learning for certain aspects. We formulate this problem as agentic evolutionary program synthesis, wherein a large language model, leveraging its pretraining knowledge and tool-use proficiency, and guided by a circuit verifier, evolves a program that defines a Clifford circuit for a wide range of curated tasks such as logical state preparation, logical gate design, and logical routines such as code-switching, code growth, etc. Our approach consistently finds fault-tolerant circuits for all of these tasks, each verified by an exhaustive check over faults.
cs.AINov 13, 2026new
Dhruv Dixit, Khushi Patel, Congcong Xu et al.
Small language models are attractive for local, low-cost, and privacy-sensitive deployment, but they remain unreliable on mathematical reasoning tasks that require decomposition, exact computation, tool use, and verification. We introduce \textbf{ARCHER}: \textbf{A}ction-trace \textbf{R}einforcement for \textbf{C}alculator-grounded \textbf{H}ierarchical \textbf{E}xecutable \textbf{R}easoning, a two-stage framework for training small language models (SLMs) to produce executable and auditable reasoning trajectories. In the first stage, ARCHER compiles free-form chain-of-thought rationales into typed action traces containing reasoning steps, calculator calls, deterministic runtime observations, verification actions, and final answers. This shifts the supervision target from imitating explanatory text to learning a tool-grounded reasoning policy. In the second stage, ARCHER refines this policy through asymmetric hierarchical reinforcement learning, using grouped and conditionally gated trajectory rewards for answer correctness, tool grounding, process discipline, and trace alignment. Across five standard sub-10B backbones, ARCHER improves over the strongest tool-augmented baseline by 5.51 percentage points on macro-average across benchmark accuracies, including gains of 8.95 points on MATH-500 and 5.33 and 5.34 points on AIME-2024 and AIME-2025. Evaluations on two additional reasoning-specialized 7B backbones further show that executable grounding remains beneficial after reasoning-focused reinforcement learning or distillation. These results demonstrate that executable supervision and reward-guided behavioral refinement provide a general and effective training paradigm for compact mathematical reasoners.
cs.MA · cs.AINov 13, 2026new
Sandip Nandi, Aishwarya Bose, Apurba Bose et al.
Autonomous trading agents need controls that prevent problematic orders without making legitimate tasks impractical. We study this aspect of agentic compliance with TradeGauntlet, a replayable limit-order-book simulator that separates detection, prevention, and execution cost. In 1,680 simulated sessions with Claude Haiku 5.5 and Sonnet 5.5, all conduct alerts were self-matches. Gates eliminated those alerts by rejecting self-crossing orders, while the aggregate-gate condition reduced Sonnet's two-sided fills by about 15%. A conduct reminder reduced Sonnet's alert rate but reduced Haiku's fills without a detectable alert benefit. These results show why zero alerts alone do not establish an agent's compliance. Scripted probes and two search methods test weaknesses beyond those observed in the LLM runs. The searches find split-order schedules that evade the implemented conduct rules; a learned baseline catches all 2,400 tested searched episodes but also flags 15% of a large maker's sessions. Aggregate caps reduce the response-dependent gain unevenly, with smaller reductions when only near-touch depth affects prices. A response fitted to five Nasdaq stocks and a separate ABIDES experiment expose the dependence of these estimates on market assumptions. A distinct adversarial prompt elicited refusals from Sonnet in all 240 rounds, yielding no candidate strategies to evaluate. TradeGauntlet provides reproducible tests of controls around trading agents; the experiments do not establish legal compliance or real-market detection accuracy.
cs.CL · cs.AINov 13, 2026new
Robin Shing Moon Chan, Ryan Cotterell
Program induction refers to the process in which a learner aims to recover a program from a few given input-output examples. This paradigm broadens programming access by letting users demonstrate the rule-based behavior they want to apply to unseen data, instead of having to implement it themselves. However, recovering a program from finite input-output examples is fundamentally ambiguous, as a finite set of input-output examples is theoretically consistent with infinitely many programs whose output diverges on unseen inputs. We view this problem from a Bayesian perspective, where, given a prior $p^0$ over programs, identifying the target program means concentrating posterior probability mass on it by observing additional input-output examples. In the interactive program induction setting, the learner chooses maximally informative inputs and queries an oracle, such as the user, for their outputs. We derive a greedy query-selection algorithm and show that it terminates after $\mathcal{O}(\frac{1}γ\log 1/p^0(f^\star))$ informative oracle queries, and returns the target program $f^\star$ up to approximate equivalence when no other class outweighs it under the prior. The algorithm uses a language model as the prior under a PAC-style relaxation of program equivalence, where the query-selection objective is estimated using importance sampling. On the CodeARC benchmark, the algorithm improves accuracy over the LM only agent by up to 22 percentage points while observing up to 10.5 fewer input-output examples.
cs.RO · cs.AINov 13, 2026new
Ivana Ragusa, Lorenzo Guiducci, Filippo Cavallo
AI-enabled remote patient monitoring is evolving beyond passive data collection toward agentic AI systems that interpret patient states, coordinate care, and initiate action across domestic and clinical settings. This structured scoping review finds that mature agentic care has not yet emerged: Most systems remain human led, while stronger decision-action coupling and embodied AI remain largely confined to prototypes and pilots. Healthcare robotics makes the transition from monitoring to action especially visible by translating AI-generated recommendations and decisions into speech, movement, prompting, physical assistance, escalation, and care coordination, making embodiment, clinical integration, human oversight, and governance fundamental design requirements. Further progress requires safe and interoperable care architectures in which intelligent components operate through clinically validated and accountable workflows. We propose agentic robotics as a future direction for supervised systems embedded in the care ecosystem, connecting AI agents and robots with clinicians, caregivers, and patients through auditable workflows that can be interrupted and remain subject to human oversight and explicit AI governance.
stat.ML · cs.LGNov 13, 2026new
Mame Diarra Fall, Éric Barat, Nicolas Dobigeon
Regular exponential-family distributions are ubiquitous in statistics due to their appealing and convenient properties. Notably, they can be expressed in terms of Bregman divergences, revealing fundamental connections between statistics, convex analysis, and information geometry. However, this geometric structure has not been systematically exploited to characterize these distributions with respect to their expectation parametrization. Building on a projective characterization, we establish a general construction of equivalent dual Bregman representations, together with normalization and uniqueness results. These representations express log-densities as ordinary Bregman divergences in dual coordinate systems, providing geometric characterizations of the parameter space. So-called $β$- and $α$-scaled dual representations emerge as particular instances of this framework. We further derive dual Bregman representations of conjugate priors and establish that they preserve conjugacy. Applications to classical distributions illustrate the geometric structures underlying seemingly different statistical models. Moreover, we show that these representations provide natural reference geometries for establishing relative smoothness of the primal Bregman representation of the distribution, with explicit bounds that are sharp in the scalar and separable settings. An illustrative application to the Poisson distribution sheds new light on several works in the literature, providing complementary geometric justifications for reference geometries while deriving existing algorithms.
cs.SE · cs.LGNov 13, 2026new
Mengfan Xu, Zirui Liu, Juan Zhai et al.
We study feedback-driven black-box processes across software engineering and machine learning. These processes are inherently sequential: policies generate inputs, observe outputs/feedback from systems, and use the feedback to guide the design of subsequent inputs. While many policies are developed under structural assumptions on the underlying systems, many real-world systems operate without such knowledge, necessitating black-box policies that lack a unified evaluation framework. We introduce a stochastic framework that models and evaluates such policies using Markov chains and simulation, thereby connecting empirical practice with theory. We instantiate the framework in mutation-based fuzzing (widely used for discovering software bugs via sequential input mutation), by a novel Markov chain model of fuzzing whose transitions are governed by the mutation strategy and whose feedback is given by program execution. We develop a rigorous evaluation paradigm with new performance measures, sound estimators, and evaluation procedures with provable guarantees. Guided by the theory, we evaluate 9 fuzzers on 18 widely used fuzzing benchmarks and conduct case studies in 4 non-fuzzing domains. These results show that the framework provides a principled basis for evaluation while connecting to practice, with implications beyond fuzzing.
cs.CL · cs.AINov 13, 2026new
Disha Sheshanarayana, Sai Linisha Reddy Goli
LLM-based verifiers are increasingly used to evaluate agent executions, but their reliability is typically measured under a single representation of the trajectory. In deployed agent systems, the same execution may be exposed through composed tools, delegation boundaries, or compacted histories, changing what execution evidence is directly observable to the verifier. It therefore remains unclear whether correct verifier judgments persist across such abstraction boundaries. We introduce Verification Transport, a reliability test that measures judgment stability under abstraction-induced changes in execution observability. We construct matched native, abstracted, and restored views of the same frozen executions while holding the task, realized trajectory, outcome, ground truth, evaluation criterion, and verifier fixed. Across four agent-evaluation benchmarks, seven verifier backbones, four verification protocols, and three abstraction boundaries, we find substantial transport failure, reaching about 31% under macro abstraction and delegation. Verification protocols exhibit distinct failure modes: direct verification more often produces incorrect verdicts, whereas evidence-grounded and deliberative protocols tend to abstain under limited observability. This suggests that lower transport failure rates do not necessarily imply greater reliability. Restoring the corresponding execution provenance recovers 85.5% of transport failures, while recovery is lower when provenance is lossy or unavailable. These results show that native-view accuracy alone is insufficient to characterize verifier reliability when execution evidence crosses abstraction boundaries.
stat.ML · cs.LGNov 13, 2026new
Christopher Marouani, Grigory Bartosh, Christian A. Naesseth
Contrastive representation learning methods operate by pulling different views of the same data sample together while pushing them away from views of different samples. A fundamental challenge is the estimation of the repulsive force, which typically requires large batch sizes, can exhibit high variance, and is susceptible to the curse of dimensionality. We introduce ScoreCLR, a contrastive representation learning method that offers a novel way to estimate this repulsion term. We consider the framework of representation learning by maximizing the mutual information between two latent views of a sample. By deriving the gradient of the Shannon entropy with respect to the encoder parameters, we obtain an equivalent objective that involves the score of the latent representation. We then use denoising score matching techniques to estimate this score with a lightweight network that is trained jointly with the encoder. We further propose regularization techniques using different noise levels and a Gaussian prior on the latent representation. ScoreCLR replaces the batch of explicit negative samples in contrastive methods with a learned repulsion whose cost is linear in the batch size rather than quadratic. On CIFAR-10 and ImageNet-100 it does so with minimal loss of linear-probe accuracy compared to SimCLR under a shared recipe, while producing projector outputs of higher effective rank, closer to the Gaussian prior.
math.NA · cs.LGNov 13, 2026new
Jianguo Hou, Wuchen Li
A diffusion equation can be rewritten as a transport equation driven by the score function of the evolving density. Conventional probability-flow formulations, which move particles deterministically with this score-driven velocity, typically conserve total mass. In this work, we develop a Lagrangian method for reaction-diffusion equations in which each particle carries a position and a physical weight. The characteristic ODEs incorporate source terms into the evolution of the log-density and its gradient, while particle weights track changes in mass. We combine this representation with a learned velocity field and weighted particle training to approximate these dynamics. For non-increasing reaction rates, a relative-entropy inequality supports the mass-normalized self-consistency residual as the training objective. This objective can control the error of the induced density when the growth rate is bounded. The algorithm divides the time interval into short windows. In each window, a frozen velocity generates the particle data for a regression of the network velocity. We draw the particles from period-adapted initial sampling distributions and weight them in the loss partly by their predicted future mass, because the reaction can make regions with little initial mass important later. We numerically investigate the method on the Ornstein--Uhlenbeck (OU) equation with a linear source term, Fisher--KPP equations, and Wasserstein--Fisher--Rao gradient flows.
cs.AINov 13, 2026new
Abhilash Durgam, Nyle Siddiqui, Mubarak Shah et al.
Surface electromyography (sEMG) from the wrists can recover typed text without a keyboard. It is usually posed as a signal-inference problem: characters are read out of the muscle signal and character error rate is reported offline. We argue that this framing removes most of the problem. The decoder is constrained simultaneously by the muscles, by the language being typed, and by typing behaviour such as typos, corrections, backspaces and line breaks, and a usable system must predict keys live while the hand keeps moving. Existing systems fall short on two counts: (1) they process sEMG in fixed windows and carry nothing from one window to the next, discarding context accumulated over a session, and (2) they lack any understanding of the linguistic structure that drives typing, so every keystroke is inferred from the muscle signal alone. We propose a fully causal, zero-lookahead sEMG typing decoder built on a state space model (SSM), in which a single model decodes a whole session without windowed partitions, conditioned on what has already been typed. We achieve this by fusing a traditional typing decoder with an efficient autoregressive language model, instilling both sEMG and linguistic understanding in one model. To train such a model for per-character typing rather than whole-word text, we introduce a retokenization method that rewrites ordinary prose as human-like typing, with keystrokes, typos, backspaces and line breaks included, converting WikiText into a large typing corpus where none existed. Our model, having never seen a single sEMG-to-character transcript in typing, achieves a state-of-the-art character error rate of 25.13% on emg2qwerty's cross-user benchmark, a 5.4% relative improvement over the previous best with 6.7 times fewer parameters, while running strictly causally with no future context.
cs.LGNov 13, 2026new
Ahmed Shoyeb Raihan, Md Habibor Rahman, Ashish Nimbarte et al.
Accurate prediction of high-entropy alloy (HEA) strength can guide composition and processing decisions; however, obtaining tensile-test data for HEAs can be costly and experimentally demanding. This study proposes Robust Outcome-aware Sequential Experimentation for High-Entropy Alloys (ROSE-HEA), a sample-efficient framework for predicting ultimate tensile strength (UTS) from alloy composition and processing conditions. Its acquisition policy anticipates how possible UTS outcomes update predictions and the support for complementary Gaussian process (GP) descriptions. It then reassesses the worst-case expected predictive risk remaining across a reference domain to select informative candidates. A retrospective evaluation on 501 HEA records compares ROSE-HEA with expected-improvement (EI) and upper-confidence-bound (UCB) Bayesian optimization (BO), random GP sampling, and eight conventional regressors across 30 matched runs with shared initial observations and test conditions. Starting from 25 initial observations and acquiring 100 additional responses, ROSE-HEA achieves the lowest mean root-mean-square error (RMSE), at 152.84 MPa, and the highest mean R^2, at 0.7443. Mean RMSE is reduced by 6.16% and 6.37% relative to EI and UCB, respectively. The selected observations improve prediction across six regressor families and provide broader design-space coverage. ROSE-HEA also surpasses the strongest conventional regressor's mean predictive performance using 125 rather than 185 observations, corresponding to 32.4% fewer training observations. These findings demonstrate the potential of outcome-aware robust sequential learning to improve UTS prediction in HEAs under constrained training budgets.
cs.DC · cs.LGNov 13, 2026new
Jianjun Ou, Matthew Lentz
Agent sessions may involve frequent context edits, which effectively invalidate cached KV state and force requests to wait for re-prefill before decoding. We present Callo, a disaggregated serving system that leverages invalidated KV for speculative drafts to reduce this latency, while verifying drafts against the exact KV after prefill before committing tokens. Callo makes effective use of decode resources by scheduling speculations during periods of underutilization and preempting speculations when drafts are likely to not be accepted. We evaluated Callo on a two-GPU testbed using SWE-agent sessions across three model families, observing all-request speedups of 1.05-1.19x over vLLM at moderate load. We demonstrate that these gains are also realized at scale through cluster-level simulations driven by real-world traces.
cs.AINov 13, 2026new
Yifan Jin, Tao Liu
Clinical world models expose language-model agents to predicted patient responses, but simulations can be unreliable for poorly supported treatments or unfamiliar hospitals. We introduce Support-Gated World-Model decision support, a lightweight extension of a propose--simulate--refine agent. A recurrent dynamics ensemble predicts short-horizon physiological responses to documented norepinephrine dose categories, while a behavior model estimates observational action support. The interface withholds predictions with insufficient support or excessive ensemble disagreement and abstains when no candidate qualifies. Evaluation on eICU-CRD includes 70 hospitals and 30,700 decision transitions, with 14 hospitals held out for testing. At 62.0% factual-transition coverage, joint gating yields a MAP mean absolute error of 7.3 mmHg, compared with 9.2 mmHg for random selection at the same coverage. The gated agent achieves 60.0% agreement with recorded actions at 81.0% decision coverage, compared with 55.0% agreement and 99.5% coverage for an ungated ensemble agent. Component ablations show lower factual risk with joint gating than with either gate alone.These results characterize a reliability--coverage trade-off in selective clinical simulation; they do not identify individualized treatment effects or establish improved patient outcomes.
cs.LG · cs.AINov 13, 2026new
Wang Ma, Debarun Bhattacharjya, Junkyu Lee et al.
Reliable use of large language models (LLMs) requires uncertainty estimates that separate ambiguity intrinsic to a prompt from genuine knowledge gaps. Bayesian and deep-ensemble methods deliver this aleatoric/epistemic decomposition in principle, but are computationally prohibitive at LLM scale and require white-box access. Existing black-box methods, in contrast, summarize to a single scalar that cannot make the distinction. We introduce a two-level ensemble that bridges this gap. Stochastic decoding of a fixed prompt forms the inner ensemble; meaning-preserving semantic perturbations of the prompt form the outer ensemble. Variability is measured in a continuous embedding space, and the law of total covariance gives an exact within-/between-perturbation decomposition to estimate aleatoric and epistemic uncertainty (AU/EU); it requires only black-box sampling access. On top of this estimator, we contribute three theoretical results that ground the framework: (i) a finite-sample bias identity for nested Monte Carlo with a closed-form bias-corrected epistemic estimator; (ii) a Bayesian bridge bounding the gap between perturbation-population AU/EU and Bayesian targets; and (iii) a separation result identifying a class of failure modes that any consistency-only black-box method provably misses while perturbation-based EU detects. Across five short QA benchmarks and four instruction-tuned LLMs spanning two families and two scales, our estimator is competitive with strong white-box baselines, and at a matched generation budget it matches leading black-box methods on error ranking while decomposing AU and EU; controlled paired interventions show that AU and EU respond selectively to distinct, actionable sources of failure: prompt ambiguity and model knowledge gaps; and a long-form pilot on the NQ long-answer split shows the framework extends to claim-level factuality without retraining.