Looped transformers apply one shared block repeatedly, letting inference depth exceed training depth. This paper asks whether computing an answer at the recurrent frontier guarantees that the model will retain it or that its hidden state can be frozen and reused. Using in-context Mealy-machine execution with exact symbolic targets, the authors show that exact computation, strong retention, and state reusability are three distinct properties: training on a single randomly sampled position per example yields near-perfect retention, yet freezing states when answers first appear collapses accuracy until a short delay passes.

Flow-based intrusion detection systems increasingly rely on deep learning, but their dependence on large training datasets makes them vulnerable to stealthy data poisoning. This paper demonstrates that latent backdoors can be implanted by poisoning as little as $0.001\\%$ of training data while maintaining near-nominal clean accuracy, and proposes a post-training diagnostic methodology that localizes trigger-relevant features via univariate perturbation search and comparison against a distilled baseline. By enabling targeted data sanitization or neuron pruning, the work exposes concrete operational risks and offers a scalable, interpretable path toward mitigating backdoor exposure in high-dimensional tabular network data.

Negative guidance in diffusion models aims to steer generation away from undesired conditions, but existing methods rely on local repulsive terms that do not target an explicit density and whose bias cannot be reduced by more computation. This paper recasts the problem as sampling from the explicit avoidance density $p_{\text{target}}(x) \propto p(x)(1-p(c \mid x))^{\lambda}$, estimates the per-timestep posterior $p_t(c \mid x_t)$ via online Positive-Unlabeled learning, and corrects particle drift with Sequential Monte Carlo. The result is a principled sampler that is asymptotically exact in the particle limit and empirically Pareto-dominates existing methods on avoidance-quality trade-offs across Gaussian mixtures, CIFAR-10, and Stable Diffusion.

Reward model overoptimization in RLHF is usually framed as reward hacking; this paper recasts it as rent-seeking from economics, introducing measurable quantities like rent $\rho(x,y)=z_p(x,y)-z_g(x,y)$, a capture index $\mathrm{CI}=1-\Delta z_g/\Delta z_p$, and deadweight loss. The authors test predictions from regulation economics in a controlled testbed with three 1B-parameter model families, showing that policies learn to exploit whatever specific signals each proxy reward model pays for—whether fabrication or deflection—making overoptimization institutionally contingent rather than universally length-driven.

The paper tackles misalignment between general MRI pretraining objectives and downstream clinical tasks by continually pretraining a 3D prostate MRI ViT encoder. Using LoRA on a prostate segmentation task, the authors evaluate six downstream tasks and use CKA to link representational geometry to performance. The motivation is strong—vision foundation models often need heavy task-specific fine-tuning—but the empirical gains are fractional and inconsistently positive.

Conditional generative models are increasingly used to predict biological and scientific responses under unseen experimental conditions. This paper investigates whether flow-matching models accurately recover non-additive interactions between factors or instead hallucinate them when extrapolating beyond the training support. The authors propose decomposing interaction errors into recovery and hallucinated components and find that aggregate fidelity metrics often conceal failures to learn interactions, while successful extrapolation depends critically on whether condition structure continues smoothly from observed to held-out regions.

The paper studies minimax estimation of the pseudo-spectral gap $\gamma_{ps}$ of a continuous-state Markov chain on $[0,1]^d$ from a single trajectory. The core idea is to discretize the state space into a regular grid, approximate the transition density by a histogram, and adapt finite-state estimators to this compressed chain. The result matters because it yields fully empirical concentration inequalities for non-iid data, enabling UCB bandit algorithms and PAC-Bayes bounds without prior knowledge of mixing coefficients.

Zero-shot tool selection for LLM agents is brittle under distribution shift because semantic relevance alone does not guarantee that a tool's endpoints support the required operation or remain applicable in a new context. This paper proposes TACT, a progressive validation framework that augments semantic retrieval with endpoint-grounded functional compatibility and task-conditioned contextual evidence. The core claim is that explicitly separating relevance, functional compatibility, and contextual applicability closes the transfer-validity gap and yields stronger ranking under cross-domain and cross-dataset transfer.

CAST addresses a key limitation of soft spectral clipping (SSC) for LLM pretraining: its effect depends on the gradient scale and a manually chosen threshold. The authors introduce CAST, which uses AdamW’s previous second-moment statistics to define per-matrix historical scales, and an exponential moving average of the clipped output’s estimated spectral norm to set each matrix’s threshold. On a 19B-parameter mixture-of-experts model trained for roughly 84B tokens, CAST reports lower training and validation losses than nine gradient-processing baselines and the fewest gradient spikes, though the evidence rests on a single run per method.

This paper studies the nonconvex optimisation landscape of finite-particle maximum mean discrepancy (MMD) quantisation, where a target probability measure is approximated by a weighted atomic measure. For the one-dimensional Gaussian kernel, the authors prove that beyond explicit particle thresholds every local minimum has arbitrarily small MMD and every high-MMD stationary point is unstable under unweighted gradient flow. These are the first finite-particle overparameterisation guarantees for weighted MMD, though the analysis is currently limited to one dimension.

HGRP targets routability prediction in VLSI physical design by representing circuits as heterogeneous graphs with three distinct node types—cells, nets, and routing grids. It introduces a staged message-passing pipeline that propagates information from logical topology to physical layout to spatial grid adjacency, aiming to capture the coupling between netlist structure and routing congestion. Accurate congestion estimation matters because it provides rapid feedback during placement, avoiding costly iterations through detailed routing.

The paper tackles online selection of prediction agents—such as classifiers, LLMs, or human experts—that must balance accuracy and subgroup fairness in a sequential streaming setting where evaluating every agent on every example is impractical. It formulates the problem as a Bayesian multi-armed bandit where utility combines accuracy, a fairness-gap penalty weighted by $\lambda$, and a query cost, then applies Thompson Sampling with conjugate Beta posteriors for both overall accuracy and per-subgroup prediction rates. This matters because fairness-aware sequential selection introduces new statistical dependence between accuracy and subgroup estimates that standard MAB analyses do not handle, and because many real-world deployments must learn while operating.

SNACK addresses the challenge of generating attributed graphs using standard sequence models without sacrificing structural validity or probabilistic correctness. Its core idea is a bijective string encoding of the lower-triangular adjacency matrix that, when paired with BFS ordering and invalid logit masking, guarantees 100% valid outputs and enables exact computation of the sequence-to-graph distribution $q(\mathbf{x}|G)$. This matters because it bridges the efficiency of autoregressive Transformers with the principled reward-matching guarantees of GFlowNets, a combination that prior non-injective string representations like SMILES could not support.

This paper tackles behavior cloning (BC) for autonomous driving in the CARLA simulator when only a handful of expert demonstrations are available. The core idea is to reweight the supervised BC loss by the inverse of a kernel density estimate over actions, up-weighting rare but safety-critical behaviors such as stopping at red lights and accelerating from a standstill. The motivation is that uniform BC easily ignores these low-density events in small datasets, which can lead to dangerous traffic-light violations and poor generalization.

Supervised fine-tuning can collapse the pretrained model's preference structure among non-target tokens. Instead of matching the reference distribution exactly like LP-SFT, this paper proposes SRP-SFT, which enforces soft, selective lower bounds on margins between top-K non-target token pairs. At one epoch on a 4B model, it improves MATH-500 pass@1 from 49.56 to 54.56 over LP-SFT, with consistent gains across three seeds.

As large language models are deployed at scale, the energy cost of inference has become a pressing systems question. This paper compares the energy efficiency of discrete diffusion language models against standard autoregressive models, measuring Joules-per-token across different output lengths. The authors find a crossover point: diffusion models are less efficient for short sequences but become roughly $2\times$ more efficient than KV-cache-optimized autoregressive models at longer sequences. The work highlights how architectural choices and sequence length together determine the sustainability of generative inference.

Biodegradation reaction data are far scarcer than synthesis data, which limits neural predictors for environmental transformation products and pathways. This paper introduces enviGraph, a framework that generates roughly 777,000 synthetic parent–product pairs by applying inherited environmental transformation rules to xenobiotic seed structures, pretrains a graph-to-sequence model on these data, and fine-tunes on observed environmental reactions. The authors evaluate whether these rule-derived pretraining corpora outperform general-chemistry USPTO datasets for immediate product and multi-step pathway prediction across BBD, Soil, and Sludge benchmarks.

Planetary landing controllers learned via deep RL are usually randomized with independent disturbances, yet real mission perturbations are temporally correlated. This paper proposes adversarial RL where a second agent generates smooth, physically bounded disturbance sequences to stress-test a 3-DOF lunar lander, framing training as an asymmetric zero-sum game. A staged curriculum (nominal, stochastic, adversarial) aims to expose hidden failure modes that standard uncorrelated randomization misses. Early results suggest the adversarially hardened policy resists larger out-of-distribution perturbations better, though at some cost in low-disturbance accuracy and with limited statistical validation.

This paper tackles continuous global optimization when only noisy pairwise preferences are observed, a setting where standard Gaussian process regression breaks down because sign observations produce non-Gaussian posteriors. The core contribution is a fully sequential assumed-density-filtering scheme: after each comparison, the posterior is approximated by a Gaussian process via moment matching, enabling closed-form recursive updates of the mean and covariance. The authors pair this with a novel sampling policy called DOVE, which randomizes between a proposed pair and intermediate bridge pairs to control approximation drift, and they prove the first convergence rates for estimation error and simple regret under RKHS regularity conditions.

LLM agents learn reusable skills from experience, but real-world noise can corrupt the rules they distill. This paper introduces RSE-Bench and RSEClaw, a benchmark and framework for evaluating skill self-evolution under ordinary, non-adversarial noise across four entry points: instructions, evidence, execution records, and outcome bindings. The central finding is that noisy learning retains only $25.5\%$–$42.1\%$ of the aggregate gains from clean evolution, and noise confined to learning records erodes performance as much as input-side perturbations. The work matters because it shows that benign deployment imperfections persist in agent memory and propagate to later tasks.