Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2604. 02995v3 Announce Type: replace-cross Abstract: We introduce the penalised Saito functional $\mathfrak S_{\lambda,\beta}(\mathcal{A};d_1,d_2)$ for a reduced arrangement $\mathcal{A}$ of $n$ lines and a prescribed pair $d_1+d_2=n-1$.
The paper investigates how multimodal large language models (MLLMs) handle conflicting evidence presented in text, image, or both forms. Across 13 MLLMs and two datasets, the authors find that models are not robust to knowledge conflict: they tend to accept contradictory image evidence more readily than contradictory text, and when both modalities conflict the preference is arbitrary, depending on input order, model, and dataset. The instability degrades multimodal retrieval-augmented generation and can be exploited by adversarial attacks, while simple mitigation techniques such as prompting, steering, and direct preference optimization largely fail, with supervised fine‑tuning offering only moderate improvement.
The paper introduces a training‑free approach to detect policy violations in large language models by treating the task as an out‑of‑distribution problem in the model’s activation space. It uses whitening‑inspired techniques to compute policy‑violation scores directly from normalized hidden activations, requiring only the policy text and a few illustrative examples. Experiments on several LLMs and policy benchmarks show the method achieves up to 86.0% F1, outperforming fine‑tuned and LLM‑as‑a‑judge baselines while being computationally lightweight.
The paper investigates item-sensitivity—whether a language model’s choice depends on the specific input—in a forced-choice signalling task derived from the board game Deception: Murder in Hong Kong. Across seven models, two families, a post‑training ablation, and three scoring rules, every tested cell shows item‑sensitivity, yet many are statistically indistinguishable from random choice and some perform worse than random. The authors term this phenomenon "consistency without alignment" and argue it undermines evaluations that rely solely on item‑sensitivity, permutation consistency, or self‑consistency without an independent reference.
The paper introduces EyeControl, an intent-driven image retouching agent that enhances visual focus by guiding attention to a target region with minimal user input. It combines a multi‑modal large language model to interpret user intent and a diffusion‑based retouching executor that aligns its attention map with a pseudo‑intent map, while an operation‑consistency constraint ensures natural global and local adjustments. The authors also present ControlArt‑Bench, a dataset for evaluating visual focus enhancement, and demonstrate that EyeControl achieves perceptually appealing results with stronger intent alignment.
Wave Function Backpropagation (WFB) introduces a wave‑parameterized learning framework where neural responses are described by amplitude, wavenumber, angular frequency, and phase, linking each observed state to its temporal interval Δt via the wave’s phase. The authors derive standard WFB gradients and a spatial‑curvature correction using the Laplacian of the wave response, and instantiate WFB in a feed‑forward trajectory predictor. Experiments show that using real intervals reduces average displacement error (ADE) by 20.4% over a baseline FFN, and that WFB remains competitive with ReLU controls, demonstrating its viability as a structured feed‑forward learning formulation.
Neural Symbolic Regression (NSR) uses neural networks as functional preconditioners to learn smooth, noise‑robust approximations of target functions in an interaction‑aware nonlinear feature space. A subsequent LASSO step extracts sparse, interpretable closed‑form expressions, while distributed hyperparameter optimization with Ray Tune and ASHA scheduling improves predictive accuracy and symbolic fidelity. Experiments on the Nguyen benchmark demonstrate that NSR outperforms SINDy and untuned neural baselines in RMSE, noise robustness, and out‑of‑distribution generalization, with ablation studies highlighting the importance of feature interactions, neural depth, and tuning strategies.
The paper introduces a hypothesis‑testing framework that embeds feature importance methods (FIMs) within a Weight of Evidence (WoE) analysis. By quantifying how strongly observed evidence supports a given hypothesis—whether from domain knowledge, ground truth, or the FIM itself—the approach evaluates FIM alignment and variability. The authors provide theoretical links between WoE and attribution variance and demonstrate the method on LIME and SHAP explanations across varied reference hypotheses.
arXiv:2609.01076v1 Announce Type: cross Abstract: Crystal generators can now propose periodic structures, but their control interfaces remain poorly matched to the mixed descriptors used in materials...
The paper introduces a retrieval‑augmented generation framework that combines fine‑tuned open‑weight language models with a hybrid dense‑lexical retrieval pipeline to synthesize corrosion knowledge. Applied to magnesium alloy corrosion, the system achieves significant improvements in token‑level accuracy and high faithfulness and context recall, while an additional Reason Map framework constructs evidence graphs to detect causal errors that standard factuality metrics miss. The modular design is positioned as a general blueprint for trustworthy AI‑assisted knowledge synthesis across engineering domains.
The paper introduces a simulation framework that uses large language model (LLM) agents conditioned on data‑driven personas to predict A/B test outcomes. These personas are built from anonymized user behavioral patterns, engagement signals, and inferred demographics, offering a more realistic population model than synthetic or rule‑based personas. The authors evaluate question design, persona data source, behavioral depth versus diversity, and population subsampling, achieving 0.75–0.90 directional accuracy on 40 real A/B tests.
VoiceLongMemEval (VLME) is a new benchmark that tests AI assistants on their ability to remember how users sounded by incorporating paralinguistic metadata—such as emotion labels, prosody descriptors, and voice events—into each conversational turn. The benchmark uses a three‑stage adversarial gate to ensure that models cannot succeed with transcript alone, revealing a significant affect gap: models gain 0.09 to 0.38 accuracy when provided with paralinguistic cues, and audio‑native models outperform standard ASR pipelines in extracting these signals. The dataset and code will be released upon acceptance.
The paper evaluates deep graph generative models against traditional network science models by comparing the topological similarity of generated networks to real-world networks and their effectiveness in identifying node immunization strategies for epidemic or misinformation spread. It finds that two deep graph generative models produce synthetic networks that closely resemble real-world structural properties, enabling them to identify effective immunization strategies.
S^3martCirc is a self‑supervised framework that jointly discovers and interprets neural circuits in large language models, rather than treating circuit discovery and functional interpretation as separate stages. It abstracts node behavior into two general computational roles that generalize across tasks and introduces a quantitative metric for assigning these roles, enabling simultaneous identification of importance and function. Experiments demonstrate that S^3martCirc outperforms existing methods in circuit discovery.
arXiv:2609.00968v1 Announce Type: new Abstract: SAR-to-EO image translation aims to generate electro-optical (EO) imagery from synthetic aperture radar (SAR) observations. Existing latent diffusion a...
arXiv:2609.00445v1 Announce Type: cross Abstract: Deployed language model safeguards (safety fine-tuning, filtering, unlearning) vary by principal only outside the model weights: filters are reconfig...
Privacy policies may contain internal contradictions in which commitments are undermined by practices documented elsewhere in the same policy. We operationalize this phenomenon, privacy washing, throu...
Large Language Models (LLMs) are increasingly deployed in interactive systems where understanding user intent precisely is paramount. A key capability for such systems is effective question clarificat...
Vision-language models (VLMs) have achieved remarkable performance by leveraging complementary information from large-scale image-text pairs. However, missing-modality inputs are commonly encountered...
Despite recent advances in large vision-language models (LVLMs), object hallucination remains a major barrier to their reliable deployment. Existing detection methods often characterize visual groundi...