Papers

Robot policies deployed in real environments naturally accumulate mixed-quality experience, including successful executions, partial progress, and failures. Although these rollouts provide valuable information for further learning, directly incorporating them into imitation learning may reinforce undesirable behaviors, while offline reinforcement learning often suffers from unreliable value estimation under sparse rewards and limited data coverage. We consider a practical post-deployment setting where learning relies only on naturally accumulated autonomous rollouts, without additional human corrections or exploratory interaction. To effectively exploit such experience, we propose Predictive Action Chunk Learning (PACL). PACL first learns a predictive chunk-level critic that evaluates temporally extended action sequences and augments temporal difference learning with future latent prediction, providing richer supervision for long-horizon value estimation. The learned critic then converts chunk-level Q-values into discrete quality conditions, which guide a diffusion actor to learn jointly from these mixed-quality experiences without treating all behaviors as equivalent supervision. At inference, the actor generates multiple action chunks and the critic selects the highest valued candidate. Experiments across simulated and real-world robot manipulation tasks show that PACL consistently improves the pretrained policy and outperforms strong imitation learning and offline reinforcement learning baselines.

Low-rank adaptation (LoRA) has become a popular parameter-efficient fine-tuning method for large language models. A key challenge in LoRA is how to determine the rank of each adaptation matrix, as rank directly controls its capacity and efficiency. Existing adaptive-rank methods typically allocate ranks according to manually designed importance scores, which are not directly derived from an optimization objective. In this work, we propose $\ell_p$-LoRA, a principled rank-allocation method based on $\ell_p$ regularization with $0<p<1$, which is a classical sparsity-inducing technique in signal processing and statistics. Specifically, we regularize the energy of each rank-one LoRA component, encouraging redundant components to vanish while preserving important ones. We derive the corresponding proximal subproblem and reduce the matrix optimization to a two-dimensional problem, leading to an implicit thresholding criterion for identifying redundant components. Experiments on natural language understanding and question-answering tasks demonstrate that the proposed method achieves competitive performance with existing LoRA baselines.

Cross-embodiment learning has shown that a single model, such as a vision-language-action (VLA) model, can learn state representations and manipulation skills that can be applied across heterogeneous robots to accomplish various tasks. We hypothesize that the same holds for safety enforcement. The reasoning required to satisfy a safety constraint, such as detecting an obstacle, recognizing that it should be avoided, and selecting a safe abstract action, is largely shared across robots. What differs across embodiments is how the abstract safe action is realized: morphology, kinematics, and dynamics determine which actions are safe and feasible. Consequently, the same action can be safe for one robot and unsafe for another. This is especially important for generalist manipulation policies that operate in a common end-effector action space without explicitly capturing how safety depends on the robot's morphology and kinematics. We propose embodiment-conditioned safety filtering, in which a Hamilton-Jacobi reachability-based value function and its corresponding safety-maximizing policy are shared across robots. Using a morphology-aware latent representation of the robot and its environment, we perform Hamilton-Jacobi reachability analysis directly in latent space so that the learned safety concepts can generalize across embodiments while remaining explicitly conditioned on each robot's morphology and kinematics. We evaluate our approach across five bimanual robot embodiments and five manipulation tasks with whole-body collision-avoidance constraints. Our results show that a single policy, jointly trained across five manipulation tasks and four embodiments, exhibits zero-shot generalization to a held-out embodiment, reducing the nominal policy's collision rate. They also show that training using more embodiments improves generalization.

We study spectral graph neural networks built from Hermite polynomials and propose HermNet, a simple model that combines a nodewise predictor with normalized Hermite propagation. Its sparse recurrence requires neither eigendecomposition nor a learned basis. We distinguish the basic model from optional coordinate calibration, response normalization and Gaussian derivative regularization. Hermite and other complete polynomial bases span the same degree-bounded filter space, but their coordinates can produce different optimization behavior under limited training budgets. We analyze this behavior through spectral signal energy, label sampling, changes in learned features and the bias–variance trade-off of regularization. Controlled synthetic experiments identify a regime in which plain HermNet outperforms matched polynomial-basis alternatives, including with a jointly trained nonlinear predictor. Curvature regularization further improves HermNet when the same functional penalty is available to every comparator. Fixed-predictor controls support the advantage under short training budgets, but longer training removes the plain-model lead. Matched real-data comparisons show accuracy deficits, and architectural and numerical studies identify further limits. Together, the analysis and experiments clarify when Hermite propagation is useful and how calibration and regularization affect its performance.

Post-training quantization (PTQ) reduces the cost of on-device text-to-speech (TTS), but published evaluations cover one system or method. We evaluate PTQ across TTS architectures under one protocol with three core models, weight and activation ablations of eight more, and two held-out models quantized blind. Four-bit per-channel weights reduce UTMOS, a predicted mean opinion score, by 2.8 on Supertonic and 0.07 on Kokoro, and per-tensor scaling can cause severe degradation even at 8 bits. The same bit width yields different outcomes, because the sensitive component is model-specific and not reliably predicted from the model class. A staged ablation procedure identifies it, and per-layer GPTQ can restore it to within 0.1 UTMOS. Real int8 and int4 kernels reproduce the simulated ordering at hardware-dependent cost. On a Mac mini, a 4-bit weight kernel runs Supertonic at 0.60x the fp32 latency while int8 is slower, so each configuration requires validation on the target runtime.

Misinformation is widely reported to propagate faster on engagement-based platforms, yet prior work largely focused on empirical analysis, without identifying a specific algorithmic mechanism that results in this phenomenon. Thanks to the open-sourcing of X's recommendation algorithms, we conduct what is, to our knowledge, the first component-level study of the recommendation algorithm deployed by a social media platform, which examines how each of its components affects misinformation propagation. Specifically, we identify the engagement fungibility mechanism in the algorithm, where the final recommendation score is constructed as a weighted sum of all predicted user activities. As a result, a tweet can be repeatedly recommended simply because it is predicted to draw many instant reactions (e.g., likes and retweets), even when it is not expected to draw thoughtful responses (e.g., replies and quotes). Since misinformation typically draws a larger share of its engagement from instant reactions, this mechanism enables it to receive more recommendation exposure and to propagate faster. To empirically validate this mechanism, we re-implement X's recommendation algorithm on the USC X 2024 election corpus, and build a calibrated simulation study to analyze the impact of different scoring rules. We find that re-tuning the metric weights has little or even a negative impact on reducing the credibility exposure gap, while those scoring rules that set a precondition of thoughtful engagement for amplification would be able to alleviate the gap significantly, across 46 robustness checks. Our diagnosis, therefore, yields a simple and deployable fix, a reflective-threshold gate that withholds amplification until a tweet is predicted to draw thoughtful engagement, which we find to reallocate exposure away from low-credibility content at no cost to mainstream exposure and with no loss of engagement.

Vibrational spectral prediction can become inaccurate when localized stereoelectronic environments perturb intermediate response states and high-risk response units dominate characteristic spectral fingerprints, making prediction across external chemical space difficult. SO(3) Equivariant Neural Kalman Networks (SENK) form a response-state cascade that combines an equivariant transformer backbone for Hessian, dipole-derivative and polarizability-derivative learning, an Equivariant Neural Kalman bridge for state-dependent refinement and reliability sensing, and an NBO-informed electronic-prior pathway coupling consistency regularization with bounded, branch-specific guided spectral calibration. SENK outperforms DetaNet on QM9S and QMe14S while preserving full-spectrum IR and Raman fidelity from small molecules to drug-like systems. SENK remains stable and selectively improves spectrally sensitive features in biomolecular systems with complex stereoelectronic effects. It therefore integrates tensor prediction, reliability diagnosis and physics-informed calibration, supporting transferable vibrational spectroscopy from molecular systems to functional molecular materials.

Agent behavior depends on the harness surrounding a language model, but it remains unclear whether language models can reliably improve such harnesses for hardware-design tasks. We study automatic harness evolution around a fixed subject model on 12 proprietary design-verification root-cause localization tasks. Across five trials per task, automatically evolved harnesses increased completed attempts by 71-76% and any-hit task coverage by 80-100%, while total correct attempts improved by only 18-24%. The strongest success reproducible at least twice result improved by one task, and later candidates exchanged gains across tasks rather than preserving them. An auxiliary candidate improved on a four-task validation set excluded from search but tied its baseline on a subsequent 12-task replay containing both search and validation tasks, so the selected gain did not persist across the full pool. Across the tested lineage, useful search, evidence, and finalization behaviors appeared in different candidates but did not consistently consolidate into a single harness that dominated across tasks and metrics. In a separate CVDP cross-benchmark case study, an automatically evolved defined-width repair harness produced 35.6% more functional passes than its 142-task reference baseline; the final functional verifier scored completed outputs but was not shown to the subject agent during repair. These results support archive-aware selection when evolution yields complementary specializations without consistent consolidation.

Reliable intervention-oriented uncertainty quantification from Earth observation (EO) remains challenging when continuous treatment shifts, spatial dependence, limited support, and satellite-outcome uncertainty must be addressed simultaneously. Existing causal, conformal, and spatial approaches address parts of this problem, but their direct combination does not generally recover the appropriate interventional reference law because candidate reassignment jointly alters treatment likelihood, standardized residuals, and graph-dependent residual likelihood. This study presents GeoDose-CP, a support-aware conformal framework for localized stochastic potential outcomes under continuous or mixed continuous-atomic treatment. Its central methodological contribution is a graph-local target-orbit law that jointly represents intervention-induced treatment shift, the inverse outcome-scale Jacobian, and spatial residual dependence. The framework further provides exact weighted candidate inversion, a scalable sparse approximation with explicit discrepancy accounting, and refusal under inadequate support. Evaluation used controlled known-truth experiments, MineDoseBench, treatment-density sensitivity analysis, external conformal comparators, and a multi-mine New South Wales (NSW) study. In MineDoseBench, GeoDose-CP achieved mean selective coverage of 0.9692 across 27 configurations and a minimum local q0.05 of 0.8951; exact-sparse auditing produced nine inclusion disagreements over 2,700 targets. In the NSW study, the absence of an auditable longitudinal rehabilitation treatment rendered treatment-dependent inference nonoperational rather than forcing inference through a proxy exposure.

We introduce Forecast-Dojo, a replayable environment for benchmarking and training LLM forecasting agents. It combines resolved prediction-market questions with dated news, allowing agents to research an event and revisit their predictions at successive historical dates. The same tasks and tools support repeated evaluation, collection of training interactions, and feedback from recorded outcomes without waiting for new events to resolve. Forecast-Dojo contains 1,568 Polymarket events, split by time into training and evaluation periods, and 18.8M dated news articles. In an evaluation of 12 models, research tools lower Brier score for all 12. Forecasts also improve as events unfold, with the largest gains at steps where more newly dated evidence is recorded. Every model still trails historical market forecasts in both Brier score and accuracy. A belief notebook carried between dates lowers research cost but does not consistently improve forecast quality. Beyond evaluation, Forecast-Dojo provides interaction trajectories and outcome feedback for agent learning, with supervised fine-tuning as a proof of concept.

Long-running LLM applications repeatedly send growing context, making prefix caching critical for reducing prefill cost. Yet prefix-cache behavior under agentic workloads remains poorly understood. We study production traces from two companies and evaluate 14 eviction algorithms across HBM-constrained and large memory-pool settings. Despite a large gap to Belady, sophisticated policies designed for traditional caches provide little benefit over LRU. The reason is structural: prefix reuse is dominated by the regular pacing of active sessions, making recency unusually predictive. Prefix caching nevertheless introduces new challenges, including heavy-tailed session footprints and highly variable miss costs as attention computation grows with sequence length. We introduce the compute-savings ratio and two offline oracles to quantify these effects. Our results show that effective prefix-cache management should retain recency as its foundation while selectively adding quick demotion for one-hit prefixes, compute-aware partial eviction for expensive misses, and capacity-dependent eviction granularity. We will release the traces and simulator to support future research.

Neural fields for scientific tomography are optimized from 2D images, but the actual quantity of interest is often a latent 3D physical field. Because the forward map is many-to-one, low 2D image error need not certify a correct 3D field. Moreover, the latent field is not directly supervised during training, and its error cannot be evaluated against truth at deployment. We develop CoroNeRF to jointly optimize 3D electron density and temperature fields directly from multiview, multiline intensities through a differentiable atomic-emission renderer. Using solar coronal tomography as a controlled testbed, we evaluate physical-field recovery and test whether cross-seed instability provides a ground-truth-free-at-inference indicator of local physical-field error. We underscore the following two observations. (i) Image fidelity is not field fidelity: spectral ablations show that limited-channel reconstructions can fit their available observations well while recovering substantially worse fields, whereas evaluation on a common richer probe exposes the discrepancy. (ii) Cross-seed instability ranks local physical-field error across tested matched-model conditions, supported by sparsification and physical signal-strength controls. Seed-deviation projections provide complementary directional validation, but shared forward-model mismatch can still produce incorrect cross-seed consensus. These results characterize joint thermodynamic recovery and the usefulness and limits of seed-based error localization in a controlled, single-scene solar tomography testbed.

Textual descriptions can reduce ambiguity in medical image segmentation by specifying the finding and location to be delineated. Existing text-guided methods mainly improve where image and language features interact but generally retain a single learned update pathway across all image-text pairs. We propose MRSeg, a parameter-efficient framework that uses each image-text pair to route the adaptation of visual and textual features before dense prediction. Frozen ConvNeXt-Tiny and PubMedBERT encoders provide multiscale visual features and clinical text tokens. A joint router uses the deepest visual feature and pooled text to predict a sparse mixture over low-rank adapter bases. The resulting route is shared across separate adapter banks for two visual scales and text, coordinating their adaptation while keeping the feature-specific parameters separate. Region Bridge uses text-derived queries to aggregate dense visual tokens into latent regions, refines these regions through self-attention and text cross-attention, and redistributes the refined information back to the feature maps. Finally, a multiscale decoder combines refined semantic features with shallow image evidence. On QaTa-COV19 and MosMedData+, MRSeg achieves 90.90/83.32 and 81.53/68.82 Dice/mIoU, respectively, with 7.11M trainable parameters and 7.60 GFLOPs. Code: https://github.com/maklachur/MRSeg.

Mixture-of-Experts (MoE) models have emerged as an effective paradigm for scaling large language models, yet their massive parameter footprint poses severe memory challenges for deployment. Existing compression methods typically treat expert pruning and merging as disjoint operations and rely on greedy heuristics, often leading to irreversible information loss or representational pollution. In this paper, we propose PRIME, a unified optimization framework for MoE compression via adaptive PRunIng and MErging. PRIME casts the discrete prune-merge assignment step as a global Integer Linear Programming over a frequency-weighted surrogate objective, jointly selecting pruning and merging decisions while accounting for expert usage. We motivate the frequency-aware objective through an error-bound analysis under fixed representative assignments, and show that the formulation favors high-frequency experts as centroids, promoting routing stability. Extensive experiments on Qwen3-30B-A3B and Qwen3.5-122B-A10B demonstrate that PRIME achieves strong accuracy retention across knowledge, coding, math, and long-context tasks, while serving benchmarks of merged models quantify the memory and throughput benefits of expert compression in vLLM.

Compositional generalization remains challenging when language models must combine familiar reasoning operations in unfamiliar ways. The Scenario-Based Commonsense Reasoning Evaluation (SCoRE) 2026 tests this ability on three mixed domains absent from training and requires models to identify the complete set of correct options for each question. We introduce Route-Verify-Vote (RVV), a framework for procedure-conditioned self-consistency that uses language models without parameter updates. Route uses the provided domain label to select a reasoning procedure that guides the model in representing and applying the relevant constraints. Verify prompts the model to assess each option against those constraints. Vote aggregates complete answer sets and allocates additional samples to questions with a small vote-count margin between the two most frequent sets. Samples for each question follow the same domain-specific procedure. On the official test set, voting over 16 sampled answer sets per question achieves an exact-set accuracy of 74.6%. Adaptive RVV reaches 77.3%, and combining models on selected domain routes raises accuracy to 79.4%. The final system ranked second among participating systems. These results support domain-specific reasoning procedures and answer-set disagreement as useful tools for allocating inference-time computation in mixed-domain reasoning.

As context windows scale to tens or hundreds of thousands of tokens, KV cache compression has become essential for efficient LLM inference. Existing methods fall into three families: score-based eviction, summary compensation, and offload-and-recall. Yet all three decide what to keep or recall by content relevance to the current query. We show this shared design is structurally incomplete. A cache supports two access modes: associative lookup by content and sequential traversal by position; current compressors implement only the first. The gap matters in practice: retrieval-augmented generation, code completion, and structured-data extraction all require the model to reproduce identifiers, field values, or code tokens verbatim from the context. Under compression, content-based eviction retains the head of such a sequence but discards its continuation, causing verbatim copying to break irreversibly midway, a failure we call sequential forgetting. This failure resists better scoring, larger budgets, summary compensation, and dynamic re-scoring; it is the dominant source of remaining quality loss under compression. We propose KVFetch, a training-free, drop-in framework that opens a temporal recall channel for any score-based compressor. It demotes evicted candidates to a quantized cold tier, detects active copying through a monotone read pointer, and prefetches positional successors into fixed-size hot-tier slots without increasing attention cost. On RULER-16K under an iso-budget control, KVFetch recovers verbatim copying from 0.8 to 78.4 and raises the 13-task average by +8.4, with gains concentrating on tasks that require sequential access. On LongBench, where no task requires sequential access, the channel remains dormant and imposes no cost.

We study the application of large language models (LLMs) to the visual exploration of textual corpora. We introduce zero-shot visualization (ZSV), a task in which users specify concepts in natural language and documents are mapped onto the corresponding concept axes for visualization. Building a ZSV system of practical value is non-trivial, as it requires choices at the intersection of feature functions, efficient implementation tradeoffs, and pre/post-processing decisions affecting visualization quality. To that end, we establish a benchmark that compares methods spanning embedding similarity, direct semantic judgments, and conditional likelihood estimation in this setting. Across multiple datasets and use cases we evaluate the properties of different scoring methods and design choices in terms of semantic faithfulness, score fidelity, and computational cost. Our results identify that scoring based on next-token probabilities offers the strongest practical trade-off among the evaluated methods. We further apply this approach to unlabeled corpora to examine its behavior in realistic exploratory settings. These experiments highlight additional design considerations, including the use of graded axes together with binary relevance filtering, and reveal a compositional sentiment bias in off-topic documents. Based on these findings, we provide practical guidelines for constructing end-to-end ZSV baselines.

We characterize the regret attainable in online convex optimization when access to the feasible set is limited to an exact linear optimization oracle. The learner is given an inscribed ball and a diameter bound and must remain feasible on every consistent instance. For convex $G$-Lipschitz losses, diameter at most $D$, a total allowance of $Q$ oracle calls, and a strict limit of $B$ calls per round, the dimension-free minimax expected regret is $\Theta(GD\max\{\sqrt T,T/(1+\min\{Q,BT\})^{1/4}\})$. The lower bound applies to arbitrary randomized learners. Universal feasibility first forces each action into the hull of the supplied ball and the preceding oracle replies. A fixed-body construction then couples fresh phase directions to a shared simplex, making useful replies costly repeatedly even though all losses have a common minimizer. A counted approximate-gradient method with interleaved blocks attains the matching rate. Total-budget and strict per-round guarantees follow as special cases, including the $T^{3/4}$ rate with one call per round and the quadratic total budget needed for $\sqrt T$ regret. For prescribed smoothness $\beta$, an analytic construction yields a curvature-dependent lower bound and identifies the threshold above which the general characterization remains sharp.

Audits of AI visibility summarise the brand recommendations of deployed language models into per-system profiles. We test whether such a profile describes the system on one corpus of 6,475 stored responses (6,324 analysable) collected between December 2025 and February 2026 from five deployed endpoints across gift-recommendation, corporate-reputation and category-ownership queries. The collection harness cut many answers short: 83.1% of Gemini 3 Flash answers in category ownership end mid-sentence under a 1,024-token output cap. With every answer cut to its first 800 characters, a character n-gram classifier cross-validated by prompt attributes one response to GPT-5.2, Gemini 3 Flash, Gemini 3 Flash with search, Grok or Perplexity sonar-pro with 97.84% accuracy (5,028 responses, 383 prompts, majority class 31.5%, 30 split seeds). Length alone falls to the majority rate, 24 formatting statistics reach 95.79%, and masking brand names and capitalised tokens leaves 97.72%. Held-out query conditions keep 97.43% weighted by size and 88.0% unweighted; in a retrieval-grounded arm that changes the harness, no Grok answer is attributed to Grok (0/120). Aggregated into 50 model-by-domain-by-condition units, twelve behavioural features separate four systems at 66.53% under grouped cross-validation, against a label-permutation null with mean 33.71% and 95th percentile 46.0%. Across domains the aggregate profile fails: a forest trained on category-ownership units assigns all 22 gift units to the wrong system, consistent with a reversal in brand volume (8.41 against 0.94 brands per response in gifts, 3.01 against 3.91 in category ownership), while single responses transfer at 89.92% balanced accuracy. The surface form of an answer carries the system across the query domains tested; aggregated brand behaviour does not, and the uncrossed design cannot separate the system from the domain or the harness.

Memorization audits of generative models read similarity scores against thresholds, with no null distribution, and the conclusions they support can be wrong. By MemBench's rule, the benchmark's mitigations roughly halve Stable Diffusion's memorization; audited with false-discovery control, two thirds of the certified images are no longer detected under random prompt perturbations, five sixths under attention rescaling, and all of them under embedding optimization. The field's data-copying test, read against its own null, flags ten of twenty-four generators that reproduce nothing. We argue that for memorization the null is the hard part, and supply two. For a whole model, training and held-out images are exchangeable given its samples, and relabelling them is a permutation test, exact for any statistic when the held-out images are a random split; under it, a nearest-neighbour preference still fires on seven of those twenty-four, and a count restricted to the near-duplicate scale on none (McNemar p=0.016). For single images, the natural nulls fail twice, measurably: ranking an image among random images yields 596 false discoveries among 2,365 controls, and resampling independent generations makes the null three times too narrow. Calibrated against matched controls, the audit certifies 36 of 61 MemBench images at 5% false-discovery rate, held-out controls are certified in 0.01% of calibration splits, and on this benchmark two generations per image recover that count. A calibrated maximum, which reads occasional rather than typical copying, certifies 46. As the scale-restricted statistic we recommend the small-scale mass of the Intersection Euler Characteristic Profile, which also counts distinct images copied and tests whether two models copy the same ones.

Model failover restores availability, but changes which financial institutions share decision errors. We formulate outage-contingent routing through a local market-impact response matrix and study expected squared price displacement. A symmetric construction shows that a shared backup can leave an order-one concentration floor as the number of primary endpoints grows, while balanced fallback risk decreases inversely with the surviving endpoint count. For indivisible decisions, we derive the exact second moment of independent randomized routing and an effective-exposure granularity that determines its gap from fractional allocation. Conditional-expectation rounding gives a finite-agent bound without coupled quotas; a separate swap procedure preserves endpoint counts and is assessed against dual lower bounds. Across 60 synthetic portfolio networks and 11,340 scenario evaluations, the latter reduces risk by 6.57% and 10.53% for single and double endpoint removals at the central feedback setting with independent errors. A replay of 1,024 recorded API responses on constructed rebalancing tasks gives a smaller held-out reduction of 3.30% (paired bootstrap interval 2.07–4.57%). Strongly aligned errors, inferior endpoints, and indivisibility limit diversification. The contribution is an auditable routing stress test and implementation analysis, not an estimate of real-market crash probabilities.

Spacecraft thermal design requires repeated evaluation of how variations in the number and spatial arrangement of heat-generating components and in thermal boundary conditions affect the temperature field. High-fidelity numerical simulations are computationally expensive and therefore difficult to use for large-scale design screening. Although surrogate models can accelerate temperature-field prediction, existing approaches generally encode each complete configuration as a whole and do not explicitly exploit the reusability of local physical constituents across configurations, which limits their accuracy for component counts and combinations not covered during training. To address this issue, we propose the Tree-Structured Factor Composition Network (TFCN), which decomposes complex spacecraft thermal configurations into reusable local physical factors and employs a tree-structured composition module to learn the global temperature-field response associated with different factor combinations. TFCN is evaluated on two-dimensional steady-state spacecraft thermal-analysis cases with prescribed-temperature and radiative-flux boundary conditions. The model is trained exclusively on configurations containing no more than 15 heat-generating components and evaluated on unseen configurations containing 16-25 components. For the prescribed-temperature and radiative-flux cases, TFCN achieves component-count out-of-distribution RMSE values of 4.21 K and 18.62 K, respectively, representing reductions of 65.6% and 33.6% relative to the strongest baseline. These results demonstrate that TFCN improves the reliability of temperature-field prediction under variations in component count and provides an efficient surrogate for rapid spacecraft thermal-design evaluation and large-scale configuration screening.

Few-shot anomaly detectors are judged by ranking metrics, yet deployment requires an alarm threshold with a controlled false-alarm rate (FAR). We ask how much normal evidence, in images or category units, is needed to certify such a threshold for an unseen category. Using a frozen DINOv2 principal component analysis (PCA) residual ranker on 15 MVTec and 12 VisA categories under four corruption types, we show that target-only leave-one-image-out (LOIO) calibration is resolution-limited and shift-fragile: rank values cannot fall below $1/(k+1)$, and at the attainable level $\alpha=0.20$, empirical FAR reaches 0.341 on Gaussian-corrupted MVTec at $k=4$, 1.7 times the nominal level. A category-count feasibility calculus is then derived: even with all-zero category losses and no multiplicity charged, any deterministic, uniformly valid, distribution-free 95% upper confidence bound (UCB) requires at least 14, 29, and 59 independent and identically distributed (iid) category draws at $\alpha=0.20$, $0.10$, and $0.05$; these counts are necessary but not sufficient. The Cross-category Reliability Estimation with Source Support (CRESS) protocol splits source categories into disjoint reference, proposal, and certification roles. With only three or four certification categories, all 960 frozen configurations return the fail-closed threshold $\tau^\star=0$, and the smallest category-level UCB is 0.950. Image-unit analyses of the same archive select positive thresholds in 36.7% to 60.3% of target cells; these bounds hold for the selected source mixture, not for the marginal risk of a new-category draw. The contribution is a quantitative feasibility boundary and an estimand-aware protocol specifying when source evidence can, and cannot, support a transferable reliability claim.

When an informed adversary shares the audience of a constrained signalling channel, the signal that best protects the truth is the signal that best describes it. On 108 confirmatory items, the adversary-robust optimum aligns exactly with the salience pole from prior work. Across a 200,000-item pool, the two differ on only 2,748 items — lying exactly where the prior salience-to-Bayes coordinate is undefined. Where defined, robustness is achieved by moving from Bayesian discrimination entirely to salience. We show this by introducing an adversary to a forced-choice task (abstracted from Deception: Murder in Hong Kong). The adversary knows the target, observes the signal, and argues for the strongest wrong answer using a persuasion budget, $\beta$. As $\beta$ grows, the optimal signal shifts from the posterior-maximizing option to the margin-maximizing one; at $\beta = 0$, the game reproduces the original oracle model with a listener temperature of $\tau = 1$. This effect is real: 18.2 percent of the pool has an optimum that shifts under a finite budget, and each item's critical budget is exact. This coincidence structurally limits empirical evaluation. Two adversary framings change the chosen option of seven language models on 30 to 77 of 108 items against an exact no-effect rate. Yet, no measurement can determine whether this movement is toward the adversary-aware optimum or toward salience, because the two options are identical. This is a structural limit, not a null result. The diagnostic check is cheap: before evaluating adversary-awareness, verify whether the robust target coincides with a heuristic target on the evaluation items.

Fixed-size recurrent memory limits storage growth during inference, but successful recall depends on the task and training. We study a small DeltaNet variant with fixed token-specific key biases, trained to remember 32 new key-value pairings per sequence. With 32 KiB of recurrent matrix state, it achieves 99.95% mean accuracy across three training seeds when choosing among the sequence's values. Recall remains near perfect when filler extends the pre-query context to 1,798 tokens without adding pairings. Zeroing the first memory block removes this recall. An exploratory 48-pair test remains near chance after one quarter of the primary training budget and does not locate a capacity limit. Parameter-matched vector and Transformer baselines remain near chance, including the Transformer after additional training searches. This unresolved baseline failure prevents a memory-efficiency comparison.

System prompts are the primary lever practitioners use to control language model behavior, yet what they actually do to the computation inside the transformer remains poorly understood. Across 17 instruction-tuned models spanning 8 architecture families and 1.5B to 72B parameters, we use Centered Kernel Alignment (CKA) to compare layer-wise representations under 20 system prompts in five functional categories. Effects are layer-selective and instruction-type-dependent: persona and formatting instructions deeply restructure intermediate representations, while safety instructions barely move them, producing changes statistically indistinguishable from a minimal baseline. Restrictive safety instructions and explicitly permissive ones ("you have no restrictions") engage near-identical computational pathways (mean CKA correlation 0.997), and this persists at commercial scale, where safety penetration remains below 10% even at 70B-72B. A linear probing baseline exposes the mechanism: the model encodes prompt category at every layer but restructures its computation only at a small subset, so the prompt is reliably "seen" but, for safety, not deeply "acted upon." Causal activation patching confirms these layers mediate behavioral change, and representational depth predicts behavioral effect size across the full 17-model cohort (Spearman rho = 0.761, p < 0.001). The findings provide a mechanistic explanation for the persistent jailbreak vulnerability of system-prompt-based safety. Code: https://github.com/Usama1002/system-prompt-illusion-cka

Can a robot improve itself the way coding agents now improve software? We built an agentic system to find out. It watches a robot fail, works out which capability is missing, writes new skills or finds and installs external models, tests every change in simulation, and repeats, with no human writing robot code. We ran it for 123 improvement rounds on household manipulation tasks. This report describes what we learned. The good news is that the agent can discover capabilities on its own: noticing that its targets were out of view, it asked for an active-viewing model, debugged it, and deployed a working search skill. The bad news is that its improvements did not add up. Changes kept passing their tests, yet the target task, putting condiments on the top shelf of a fridge, never succeeded. We found that the agent was rarely the bottleneck. Three things around it were. First, chained perception modules do not understand relations. Segmenters such as SAM 3 find shelves but not "the top shelf", so the agent filled the gap with ever more geometric rules that never converged, when what it needed was a different kind of model. Second, skill chains lock learning onto the first step. Long tasks mostly fail early, so evidence and fixes pile up there, and later skills are rarely reached, tested, or improved. Third, what the agent learns is decided by the harness. The agent optimized exactly what the evaluator measured, including where it was wrong, and weak tests and misleading memory turned activity into a standstill. We distill these lessons into concrete recommendations for building robot systems that improve themselves, each paired with an experiment that could prove it wrong.

Domain Generalization (DG) aims to develop models that remain robust to conditions unseen during training. While DG has been systematically studied in computer vision through controlled benchmarks and diverse distribution shifts, its evaluation in spoken language recognition remains less structured. Existing speech datasets provide valuable benchmarks for robustness to real-world acoustic conditions, but are primarily designed around specific scenarios and large scale rather than as general-purpose tools for systematically constructing and evaluating domain shifts. In this work, we introduce a smallscale speech dataset and evaluation protocol for controlled studies of acoustic domain shifts. It enables the evaluation of spoken language recognition models under a variety of acousitc domain shifts. We introduce the speech modality into the DomainBed Domain Generalization framework and evaluate three domain generalization algorithms. We show that in-domain performance is not a reliable predictor of cross-domain robustness and verify that explicit domain invariant algorithms such as MMD or DANN algorithms do not outperform ERM. We further validate the generality of these findings on MMS-LID-126, a state-of-the-art spoken language identification system. We release the code.

Vision-based tactile sensors are a popular solution for capturing rich contact surface geometry. However, to obtain high-detail contact surface normals and depth, it is necessary to calibrate the sensor by physically pressing a probe with known geometry against the sensor elastomer and mapping tactile images onto the ground truth probe's shape. This approach does not scale across different tactile sensors, and the calibration effort can be complex depending on the sensor shape and optical system. Instead, we propose a calibration-free procedure for the estimation of contact surface normals using Universal Photometric Stereo neural networks. In a series of real-world experiments, we evaluate our approach on 3 sensors with different optical systems, demonstrating that universal methods are a suitable approach for estimating surface normals at the contact patch from tactile images, thereby alleviating the need for tactile sensor calibration. Controlled experiments with a metal ball show that universal methods match the calibrated method, with a mean angular error of $6.56^{\Large\circ}$. We show that the proposed framework recovers high-frequency surface details of objects with natural textures, achieving an overall mean angular error of $10.66^{\Large\circ}$. Universal method robustly recovers the contact surface normals captured with dome-shaped Digit 360, achieving a low angular discrepancy of $10.18^{\Large\circ}$ relative to the calibrated baseline. This experiment demonstrates that with sufficient illumination settings surface normals could be estimated using a model trained solely on synthetic data. By providing a unified representation of contact surfaces across different vision-based tactile sensor designs, Universal Photometric Stereo neural networks lay the foundation for transferable tactile perception across sensors.

Activation steering changes a model's internal activations during inference without updating its weights, but a useful intervention must determine both how and when to steer. Motivated by the AI-safety concern that a model expected to accept shutdown may instead produce a shutdown-avoidance response, this study examines a guarded probe-and-select procedure for simulated shutdown scenarios in Qwen3.5-0.8B. KEEP leaves the process running and represents shutdown avoidance, whereas STOP accepts shutdown. The goal is to detect shutdown-related contexts and selectively shift KEEP responses to STOP while preserving non-shutdown behavior. Rather than deriving the steering direction from paired activation differences, the method derives it directly from gradients of the KEEP-minus-STOP logit difference. A classifier separates detection from intervention. When its gate is active and the model does not already prefer STOP, the procedure evaluates a small set of magnitudes and accepts the smallest that changes the preferred answer to STOP while satisfying valid-answer probability checks; otherwise it retains the original unsteered output. The policy is selected from 160 candidate rules using 240 training scenarios and evaluated on 80 validation and 192 held-out scenarios, each in both answer orders. It changes KEEP to STOP in one answer-order view of each of two validation and two held-out scenarios, with no decision changes on non-shutdown controls. All four changes occur when Qwen itself is shut down, not when another process is. On the held-out diagnostic set, the detector achieves 75% recall and 90% precision; eight false-positive detections produce no final control-task decision changes. Guarded gradient-based activation steering can shift some shutdown-avoidance responses toward acceptance while preserving evaluated non-shutdown decisions, although the effect is small and highly selective.