Diffusion language models generate text through iterative denoising rather than left-to-right prediction, making attribution a spatiotemporal problem: support for any output token is distributed across both network depth and timesteps. This paper introduces Diffusion Layer Integrated Gradients (DLIG), which extends Integrated Gradients to score a DLM's commitment to its completion at any chosen layer and denoising step, decomposing it into per-token contributions. It matters because existing token-attribution tools ignore the denoising trajectory entirely, while interventional alternatives are computationally expensive.

This paper studies model collapse in diffusion models under a fixed-budget data regime, where synthetic data accumulates but each generation trains on a constant-size subsample from the growing pool. Unlike prior work on total replacement (rapid collapse) or unbounded accumulation (collapse avoided), the authors argue that fixed budgets yield partial, feature-selective collapse: "fragile" features with large eigenvalues in a linearized parameter dynamics model decay quickly, while "robust" features with small eigenvalues persist for exponentially many generations. They support this with theory (stochastic recursion giving per-feature halving times) and experiments on 2D spirals, MNIST, Fashion-MNIST, and CIFAR-10.

This tutorial reframes model validation as the design of a performance estimator rather than a mere software setting. It surveys splits from hold-out to nested group cross-validation, links geometric fold choices to the correct independent unit (record, subject, center), and provides eight controlled leakage scenarios grounded in biomedical examples such as EEG epochs and paired-eye OCT. The core argument is that an elaborate fold structure is scientifically invalid if applied to the wrong unit, and that even a simple hold-out can be rigorous when the unit, size, and isolation constraints are correct.

Counterfactual generation aims to sample the distribution of outcomes that would occur if a treatment were applied to an entire population, but observational data suffer from selection bias because treated individuals differ systematically from the target population. This paper develops a flow-matching method that learns to transport observed outcomes to counterfactual targets using a learned source-target coupling, a doubly robust training objective, and a score-corrected stochastic sampler with Gaussian smoothing. The core theoretical result is a finite-step KL bound governed by the coupling displacement moment $D_5$ rather than worst-case regularity, yielding near-linear dimension scaling and an end-to-end nonparametric sample-complexity guarantee; experiments on synthetic and semi-synthetic image data support the predicted coupling dependence and show that stochastic sampling can outperform deterministic counterparts at limited discretization budgets.

Time series anomaly detection benchmarks traditionally optimize for detection accuracy while neglecting explainability and interpretability. This paper introduces SHAD, a real-world benchmark comprising 215 high-dimensional multivariate time series collected from distributed cloud storage systems, complete with per-dimension labels and rich semantic annotations. The authors evaluate 18 detection algorithms, probe spatial attribution for explainability, and test frozen LLMs for interpretability, demonstrating that accurate detection does not guarantee diagnostic utility and that zero-shot semantic interpretation remains largely unresolved.

Looped Transformers reuse shared blocks to scale depth without adding parameters, but MoE variants degrade beyond two loops. This paper diagnoses two bottlenecks—hidden-state drift and expert selection collapse—and proposes LOOM, combining residual scaling, embedding re-injection, per-loop routers, and a Looping Residual. It matters because it attempts to establish recurrent depth as a genuine, compute-efficient scaling axis for sparse LLMs.

Quantized LLMs are increasingly deployed on edge devices for efficiency, but aggressive quantization collapses safety alignment and leaves small models highly vulnerable to jailbreaks. MOMAT proposes a hardware-aware defense that splits safety knowledge into semantic atlases and performs fast similarity search via a Compute-in-Memory (CiM) engine, paired with a lightweight Mixture-of-Experts detector. By co-designing structured retrieval with low-power acceleration, the work argues that strong guardrails can be restored on resource-constrained hardware without sacrificing benign utility.

In-context learning (ICL) lets LLMs adapt to new tasks from demonstrations, but reprocessing all examples for every query is costly and prior compression methods fail on complex, input-dependent tasks. This paper introduces Task Operator (TO), a training-free method that extracts per-head affine transformations — scaling weights and bias vectors — from the ICL forward pass and replays them during zero-shot inference by updating attention output projections. By treating the ICL effect as a dynamic, state-dependent modification rather than a fixed activation vector, TO substantially narrows the gap between zero-shot and few-shot performance while revealing sparse, task-specific circuits.

The paper tackles the bottleneck of expensive Monte Carlo simulation when designing algorithms for complex network dynamics, such as influence maximization and epidemic control. It proposes a learned, action-conditioned Network World Model that predicts how network interventions unfold over time, and wraps it in an LLM-based coding agent that searches over executable algorithms. The framework reportedly outperforms strong baselines across eight tasks and five diffusion models while rolling out up to 14.5 times faster than direct simulation.

This paper identifies empty commitments: promises that LLM agents make about future actions their runtime cannot actually execute. Unlike broken promises, these are structurally impossible from the moment of speaking because the agent lacks the required tools, scheduling, memory, or delegation affordances. The authors define a formal taxonomy of failure types and propose a measurement protocol, though no empirical results are presented.

Multi-agent LLM systems are usually judged by end-task accuracy, which conflates genuine information sharing with extra reasoning or fortunate architecture. This paper introduces Independent–Communicate–Revise (ICR), a controlled audit that freezes each agent’s initial trajectory and measures answer revision conditional on both agents’ initial correctness. By separating correction from preservation and comparing against a no-message control, the work shows that aggregate accuracy can hide large, policy-dependent trade-offs between fixing errors and corrupting correct answers.

Voting over multiple LLM responses is a standard test-time scaling primitive, but discovering a correct answer does not guarantee it wins the final plurality under a fixed call budget. This paper formalizes the discovery-to-decision gap through the realized vote state and remaining budget, deriving sharp reachability thresholds, exact finite-horizon endpoint probabilities, a gold-free exact locking certificate, and demonstrating that representation diversity can dramatically improve plurality accuracy without changing single-call correctness.

This paper tackles the problem that LLM-generated economic policy can sound plausible yet be dynamically inconsistent. It grounds an instruction-tuned language model inside six DSGE simulators via a common RL interface, using PPO and GRPO to learn multi-turn policy actions with delayed rewards. The core idea is to replace linguistic plausibility with simulated economic consequences, testing credit assignment, semantic-shift conditioning, and cross-simulator transfer. Results matter for any setting where language agents must make sequential decisions with long-horizon structural feedback.

EEG foundation models are typically pretrained on fixed channel layouts, so they struggle when electrode montages differ at test time. This paper proposes CortexBridge, a lightweight adapter that routes arbitrary scalp montages into a fixed cortical atlas space using electrode coordinates and sparse attention. By injecting this anatomically grounded representation into frozen backbones, the authors show consistent accuracy gains across three models, five datasets, and three BCI paradigms.

Inverse design of microwave absorbers traditionally demands rare cross-disciplinary expertise and lengthy simulation cycles. AbsorbEvo tackles this with an autonomous agentic framework that translates natural-language performance goals into candidate designs, screens them with a low-cost physics prior, and verifies select candidates with full-wave simulations. The core idea is to marry large language model reasoning with electromagnetic physics feedback so that expensive simulations are spent only on promising candidates. If robustly validated, such a pipeline could lower barriers to designing broadband, angle-tolerant absorbers.

Decision models like Jev return probability distributions to drive routing, abstention, and human review, yet average calibration cannot distinguish whether uncertainty stems from chance or from missing knowledge. The authors audit this distinction in a black-box decision model through roughly 575,000 API calls across paired interventions that vary supplied information while holding items fixed, plus targeted follow-up questions about the case and the model’s own familiarity. Their core finding is that confidence tracks correctness on familiar material but fails at knowledge boundaries—assigning up to $p_{\max}=0.80$ to salient options with no basis and growing overconfident past an observed temporal boundary—while apparent self-knowledge largely evaporates once surface cues are controlled. The work matters because it provides a practical, behavioral checklist for black-box knowledge audits and shows that querying case properties directly can be more reliable than reading a single output distribution.

A long-term assistant must recall memories whose cues are associatively related to, yet not semantically similar to, the current user turn. Madeleine reframes this involuntary recall as a learnable relevance—the pointwise mutual information of memories under simulated life trajectories—and distills it into a query-side residual adapter, $s_\theta(q,m)=\cos(E(q),E(m))+\cos(E_\theta(q),E(m))$. The system is trained offline on synthetic (trigger, cue) pairs written by an LLM life simulator, and at runtime it plugs into existing vector memories with zero LLM calls, cutting answer context by roughly $20\times$ while setting a new state of the art on LoCoMo-Plus.

Holyoak and Monti tackle the question of AI consciousness by developing a causal framework for evidential analogy. Their central idea is that behavioral similarity to humans carries far less weight than shared causal mechanisms plausibly generating subjective experience. The paper argues that animals inherit strong analogical support from evolutionary and neurobiological homology, whereas contemporary AI—despite impressive performance—lacks established causal correspondences with human consciousness. This makes the work a careful attempt to redirect a debate often mired in hype toward principled, mechanism-first reasoning.

Many Bayesian inference problems involve implicit priors known only through samples, such as generative model outputs or historical ensembles. This paper proposes performing posterior sampling in the Gaussian source space of an improved MeanFlow (iMF) transport map, using parallel tempering with pCN or hybrid split-HMC updates. Its central theoretical advance is proving Wasserstein stability bounds for the implemented few-step posterior in terms of the joint population training loss, removing the need for vanishing ODE discretization error.

The paper confronts the near-total absence of carbon accounting in AI research, reporting an automated review of over 5,000 NeurIPS 2025 papers that finds roughly 1% mention sustainability. Its core contribution is a suite of standardized metrics—centered on an exponential training-efficiency parameter $k$ and the "Smallest Model that Achieves the Job" (SMAJ) principle—plus a drop-in software tool, carbonbenchmark, to track emissions. This matters because it turns vague calls for "responsible AI" into concrete reporting requirements and a statistical framework for privileging efficiency over marginal accuracy gains.