KV cache compression for multimodal LLMs usually evicts visual tokens once at prefill, but this one-shot selection breaks down during long-form generation because the model's attention drifts as the description unfolds. The paper proposes Rate-Controlled Refresh (RCR), which uses a low-rank drift detector to decide when and where to refresh the cached tokens, while a per-sample rate controller decides how often, ensuring low-drift samples are not starved of refreshes. The core claim is that a single configuration transfers across models and tasks, outperforming per-step retrieval baselines at comparable CPU-GPU traffic.

Privacy auditing without retraining is essential for auditing deployed or third-party models, but available non-members often come from a different distribution than the training data, causing standard membership-inference attacks to conflate membership with domain shift and severely overestimate leakage. This paper proposes shift-aware score transport: the auditor learns how membership-inference scores depend on pretrained data representations from the available out-of-distribution non-members, transports that conditional relationship to the in-distribution covariate space implied by known members, and then corrects the estimate via controlled internal calibration tasks. The core contribution is a practical way to recover stable, relatively tight empirical lower bounds on $\mu$-GDP when the auditor cannot retrain, insert canaries, or control data inclusion.

Nepali–English code-switched ASR suffers from script confusion and switch-point fragility, yet human annotated conversational audio is costly. This paper curates a 15.45-hour spontaneous podcast benchmark and adapts a 1.2B-parameter multilingual encoder–decoder model using only 0.95 hours of synthetic dual-engine TTS data. The core finding is that supervised fine-tuning drives almost all cross-domain transfer, while subsequent GRPO and script-aware rewards add only marginal, hard-to-disentangle gains. The released data and the cautionary ablation make the work practically useful.

Regulated institutions must produce technical evidence for LLM deployments under the EU AI Act, yet existing benchmark suites are engineer-facing command-line tools while accountable reviewers are non-technical managers. VERA is an open-source, self-hosted dashboard that closes this workflow gap by letting reviewers select a model, run public benchmarks, and read requirement-level scores—with decomposable evidence and run-level energy consumption—in a browser. Its core contribution is wrapping raw measurement into a review and filing process designed for institutional approval.

Conformal prediction for non-Euclidean responses is underdeveloped because conventional scores require residuals or densities that assume a vector-space structure. This paper introduces the conditional expected local radius, $R_\beta(\omega|x)=\frac{1}{\beta}\int_0^\beta Q_x(\omega,\tau)\,d\tau$, as a nonconformity measure in general metric spaces, and estimates the conditional distance profile, the CELR, and its conditional distribution from a single metric distributional random forest. The result is a computationally cheaper alternative to the conditional profile score method of Zhou and Müller (2025) with uniform consistency and asymptotic conditional validity guarantees.

Coding-agent benchmarks increasingly grade agents against upstream test suites that are paired with natural-language specifications, yet those tests often require unstated names, module paths, or literals the specification never gives. This paper formalizes that mismatch as the specification–oracle gap and implements a fast, model-free static audit for five languages, then shows that on SWE-Milestone roughly a quarter of graded milestones contain hard hidden dependencies that can block entire build units. The finding matters because it reveals that many benchmarks inadvertently measure whether agents reproduce upstream implementation details rather than whether they satisfy the stated requirements.

Diffusion large language models introduce a bidirectional infilling mechanism that poses a distinct jailbreak surface compared to autoregressive generation. This paper proposes RMT-Jailbreak, which first elicits a harmful response from a source autoregressive model using existing jailbreak methods, then converts that response into a sparsely masked template and feeds it to a target diffusion model to reconstruct the harmful content. The key finding is that query-specific harmful responses provide more effective reconstruction cues than generic attack scaffolds, yielding strong cross-paradigm transferability.

The paper tests whether open LLMs that a hospital could self-host are ready for clinical use, evaluating four models (8–27B parameters) on the Spanish MIR residency exam under its real scoring and ranking rules. The authors go far beyond headline accuracy to probe calibration, the value of abstention under a wrong-answer penalty, the soundness of clinical reasoning, and whether vision models actually use the images they receive. Their headline message is that a model can rank high enough to choose almost any specialty while remaining overconfident, correct for the wrong reasons, and blind to the clinical plates it is shown.

QuaNAS tackles query-efficient neural architecture search by treating a library of earlier searches as explicit but uncertain knowledge. Its core idea is to encode design decisions as additive main-effect priors that seed a Gaussian process, letting every new evaluation re-weight trust in transferred knowledge and fall back on zero-cost proxies when the library is uninformative. The paper tests this under a frozen, self-preregistered protocol on 12 held-out datasets spanning images, time series, text, and tabular data.

Complex reasoning training hinges on two supervision modes—sparse RL rewards and dense distillation—but combining them naively causes optimization conflicts.
GIFT introduces a token-level geometric filter that decomposes the teacher gradient into a conflict-measuring parallel component ($a_i$) and a novel-guidance orthogonal component ($b_i$), then progressively anneals which tokens receive distillation.
This yields consistent improvements on hard math and logic benchmarks, with relative gains reaching 9.5% on the most difficult tasks.

SABLE tackles expensive optimisation in the high-dimensional latent spaces of pretrained generative models — a regime where latent dimension $D$ can reach $40{,}500$ while the evaluation budget $N$ is orders of magnitude smaller. The core idea is to use a Bayesian affine surrogate on the characteristic-radius sphere, decoupling acquisition into a data-sized in-span component and a scalar off-span norm that is prescribed by a finite-horizon schedule rather than left to expected improvement. Together with a fast recompression scheme that bounds model-refitting cost to $\mathcal{O}(M^{2}D+M^{3})$, the method achieves tens-of-millisecond decisions and strong sample efficiency on synthetic and scientific benchmarks.

This paper studies how to handle aging textual signals in weekly portfolio allocation: keep stale news with explicit timing information, or filter out reports older than a fixed threshold. A frozen LLM extracts ternary $z_{i,t} \in \{-1,0,1\}$ signals from SEC disclosures, then either a Double DQN RL policy or the same LLM selects portfolios for a ten-stock US universe. The core finding is that filtering old signals outperforms retaining them with age metadata for the LLM agent, yet nearly none of the return differences are statistically significant. The work matters because trading costs make signal timeliness practically relevant, but its claims of superiority rest on noisy, short-sample evidence.

Existing AI evaluations treat economic value and energy consumption separately, leaving a critical gap in understanding whether more capable systems actually deliver better returns per watt. This paper formalizes Net Value per Joule ($NVJ$) and Time Saving per Joule ($TSJ$), pairing judged professional output quality with end-to-end energy measurement across 220 real-world tasks. The work matters because it shows rankings of occupations and models shift dramatically once energy is accounted for, challenging the assumption that raw capability or dollar cost alone guide efficient deployment.

Different XAI methods frequently contradict one another, yet most tools show only one explanation at a time. XALIGN is a visual analytics workflow that places rule-based explanations and three feature attribution methods side by side for a single Random Forest model. Its goal is to let domain experts spot agreements and contradictions, calibrating trust rather than accepting a single explainer's story.

Tool selection across corpora of tens of thousands of APIs is a critical bottleneck for LLM agents, because standard retrievers rank by relevance without knowing which tools a multi-step instruction actually requires. This paper post-trains small open Qwen3 models (1.7B to 8B parameters) via supervised fine-tuning and reinforcement learning to decompose a request and select tools through a lightweight Plan–Retrieve–Select workflow. The headline result is that an 8B checkpoint outperforms a 402B-parameter Llama-4-Maverick on ToolBench and transfers zero-shot to unseen tool domains and retrievers, suggesting that tool selection can be compressed into a small, locally hostable policy rather than delegated to massive proprietary models.

Formal verification of neural networks usually requires expensive set propagation whose error accumulates with depth and input size. This work introduces a mixed-monotone architecture built as the difference of two monotone subnetworks, reducing tight output enclosure to just two forward passes at opposite interval corners. Because the enclosure size is a closed-form differentiable function of the weights, it can be minimized directly during training, yielding orders-of-magnitude speedups and near-optimal tightness without sacrificing predictive accuracy.

This paper tackles the underexplored problem of token sparsification in audio transformers, where self-attention cost grows quadratically with input length. The authors systematically compare pruning and merging strategies across model sizes and datasets, and propose frequency-aware heuristics—Spectral Flatness (SF) and Low-Frequency Energy (LFE)—to score token importance. Their core finding is that audio-specific spectral cues consistently outperform generic attention-only methods in post-training settings, offering a practical path to faster inference on spectrogram-based models.

The paper asks who adapts to whom when a confident LLM is wrong. In a within-subjects experiment, 32 participants learned science topics from GPT-5-mini assistants whose briefings were covertly falsified at 0%, 25%, 50%, 75% or 100%. The study finds that learners absorbed roughly two-thirds of stated falsehoods regardless of overall assistant reliability, while their confidence and trust ratings failed to track accuracy. By reframing trustworthiness as a property of the human-AI pair and introducing a Human Deference Index, the work argues that sycophancy evaluation must measure human-side deference as well as model-side compliance.

This paper studies whether width and depth limits commute not just at initialization, but during training, for linear residual networks with $1/\sqrt{L}$ branch scaling and $\mu$P width scaling. Its core result is that under clipped gradient descent, the trained output converges in $\mathbb{L}_2$ as $\min\{n,L\}\to\infty$ for any fixed number of steps, proving the order of limits does not matter. Without clipping, commutativity holds only in probability, because the second moment can diverge with depth even after a single gradient step. The finding matters because it rigorously characterizes joint scaling limits in a tractable deep-learning model and isolates gradient clipping as a key enabler of stable infinite-depth training.

GraphRAG usually assumes a centralized knowledge graph, which clashes with real-world settings where sensitive data is siloed across organizations that cannot share raw graphs due to privacy or ownership constraints. This paper proposes FedGRAG, a federated framework that routes queries to relevant clients using lightweight semantic prototypes and then reconstructs abstracted evidence subgraphs on a central server to enable cross-client reasoning. The problem is practically important because legal and competitive barriers prevent the global construction and indexing required by standard GraphRAG pipelines.