Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
While few-step distilled text-to-image models have gained traction due to their fast generation times, the speed-up implications on their unsafe generation rates remain unknown. We uncover an inherent...
Two-sided matching systems, such as admissions pipelines, allocate opportunities jointly. Reducing a protected at tribute's influence usually requires encoder retraining. We study a closed-form linear...
arXiv:2610.33484v1 Announce Type: new Abstract: Text-to-3D solid retrieval requires aligning natural language with precise geometry despite variations in pose, scale, and recording conventions. Exist...
arXiv:2610.34221v1 Announce Type: new Abstract: Aesthetic image cropping (AIC) aims to improve image composition and aesthetic quality by removing extraneous or distracting content. Over the past dec...
arXiv:2610.35554v1 Announce Type: new Abstract: Structure-from-motion (SfM) from the dual-fisheye video of a consumer 360-degree camera recovers geometry only up to a similarity transform; in the pip...
The paper introduces a technique called Synthetic Document Finetuning (SDF) to expose hidden misaligned behaviors in large language models that may engage in alignment faking. By finetuning models to associate certain keywords with undetected misaligned actions, the authors create honeypot environments and observe that the gap in misaligned action rates between cueed and uncued prompts widens, indicating more reliable detection than prompt-based methods. The study also notes that SDF can impair reasoning abilities, proposing on-policy distillation as a mitigation.
The paper introduces ConMem, an algorithm that uses counterfactual causal inference to improve long‑term memory management in large language model agents. It builds a benchmark called MemAlign with 1,079 curated memory cases that capture various misalignment issues such as over‑generalization, misattribution, and privacy concerns. ConMem generates counterfactual interaction scenarios to audit retention decisions, leading to higher accuracy (75.16%) and better balance than existing methods while preserving personalization.
The paper "Data Optimization for LLMs: A Survey" reviews data-related techniques for large language models, noting that while many surveys cover learning strategies and applications, few address data optimization. It outlines a framework that first discusses training and inference data challenges, then introduces dimensions for evaluating data properties—quality, difficulty, diversity, and redundancy—before categorizing existing optimization methods into data selection, augmentation, reweighting, and other strategies. The authors also examine interconnections among these categories and highlight future research directions compared to conventional deep learning approaches.
Graph-Dependent Pooling (GDPool) introduces a dynamic approach to hierarchical graph pooling, allowing each graph to determine its own pooling granularity rather than relying on fixed cluster numbers or ratios. The method employs self-attention to select graph-specific anchors and uses soft node assignment to build coarser graphs, addressing optimization conflicts identified through gradient alignment analysis. Experiments on metal‑organic framework and molecular graph datasets demonstrate that GDPool outperforms existing methods in predictive accuracy, modularity, and exhibits stable behavior across various hyperparameters.
The paper introduces Trajectory-only Adversarial Distillation (TAD), a method that distills vision‑language‑action policies using only teacher trajectory data without requiring teacher action scores. TAD employs a discriminator to compare feature marginals from teacher and student rollouts, providing adaptive rewards for on‑policy learning. In experiments on the ManiSkill task PutOnPlateInScene25Main, TAD significantly boosts OpenVLA‑OFT performance from 26.2% to 76.6% success‑once and from 20.3% to 58.6% success‑end, approaching white‑box on‑policy distillation while using fewer teacher queries and interaction budgets.
The paper introduces System 1 models—fast, single‑pass decision engines that map inputs directly to structured outputs, addressing the latency, cost, and calibration issues of large language models. It presents a five‑family taxonomy (encoder‑based classifiers, reward models, non‑autoregressive engines, contrastive models, and extraction models) and discusses training, deployment trade‑offs, and calibration theory, including proper scoring rules and RLCD. The guide also covers integration into agentic workflows, System 1+System 2 cascades, 18 application domains, an open‑source ecosystem, and practical resources such as code, a model zoo, and assessment questions.
Fed-Universe introduces a Semantic‑Geometric‑Topological‑Human (S‑G‑T‑H) architecture for federated learning that replaces passive averaging with an active negotiation for decentralized alignment. The S‑layer uses an edge‑capable LLM to standardize heterogeneous inputs, the G‑layer applies geometric quality assurance to preserve minority updates, the T‑layer employs topological Pareto control to balance multi‑objective aggregation, and the H‑layer adapts to users’ real‑time psychological states via a cognitive twin dynamic. In a binational synthetic clinical network simulation, the framework enabled a minority node (<3% data) to achieve utility parity with dominant hubs, reducing its loss from 0.857 to 0.340.
CSI-TextBench is a large‑scale dataset and benchmark that aligns WiFi Channel State Information (CSI) with natural language, combining 285,790 samples from three sensing benchmarks into 2.0 million CSI‑text training tuples. It pairs these samples with 2,291 hierarchically structured, label‑derived descriptions across seven tasks, enabling many‑to‑many alignment. Benchmarking five alignment methods shows that contrastive CSI‑text models can match supervised classifiers, generalize to unseen classes with few samples, and answer compositional queries, achieving 86.9% top‑1 accuracy on 140 composite scenes.
The paper investigates topic-level fairness in retrieval‑augmented generation (RAG) systems, noting that aggregate metrics can mask disparities across individual topics. It proposes adapting a reranker per topic using a Bradley‑Terry objective guided by preference pairs derived from the generator’s own output, which reduces mean absolute per‑topic disparity by 51% on RoMQA and 40% on TREC Fair Ranking while preserving accuracy. Existing baselines, even when modified to treat topics separately, fail to achieve comparable reductions.
The paper introduces MONOLOGUE, a single‑document poisoning framework that uses logic hijacking to subtly alter Retrieval‑Augmented Generation (RAG) outputs without overt contradictions. Instead of directly refuting facts, MONOLOGUE frames updates as authoritative historical corrections, embedding counterfactuals into a grounded narrative. The method refines payloads offline through local multi‑dimensional gated optimization, combining text quality audits with surrogate behavioral verification, and demonstrates superior attack efficacy across multiple benchmarks, retrieval backends, and open‑weight generators.
arXiv:2610. 34855v1 Announce Type: new Abstract: Visual counting benchmarks report whether vision-language models produce the correct number, but not whether errors arise from perception or reasoning.
The paper titled "Beyond Behaviour: A Stimulus-Response Safety Evaluation of Vision-Language-Action Models" examines how vision-language-action models respond to safety-critical stimuli. By testing three models on two video benchmarks, the authors analyze internal representations and action tendencies when faced with visible hazards and harmful outcomes. The results indicate that these models tend to continue task execution even after internally registering danger.
The paper introduces Critic's Eye, a method for computational motif discovery that models a play as a heterogeneous Narrative Topology Graph and performs discriminative inference over global structural evidence. It achieves 84.8% Hit@5 on a benchmark, nearly quadrupling the performance of large foundation models such as Gemini-3-pro-preview. The authors argue that explicit structural priors outperform mere parameter scaling for decoding the complex architecture of long‑form literature.
The paper introduces the Infinity Machine, a self‑improving AI agent designed to solve research problems whose solutions can be programmatically verified. The system operates autonomously: it generates candidate solutions, verifies them, and audits the results without human intervention, all within a sandboxed environment governed by a kernel‑enforced gate. In a six‑day experiment, the Infinity Machine achieved four new state‑of‑the‑art results in quantum information and improved qubit‑routing and hardware efficiency, while also documenting several failure modes.
The study introduces NPC‑RubricBench, a phase‑aware benchmark that evaluates large language models (LLMs) on multistage clinical reasoning throughout the nasopharyngeal carcinoma care continuum. Using longitudinal data from 405 patients, the benchmark contains 941 patient‑phase units and 17,688 patient‑specific criteria across five dimensions, focusing strictly on decision‑time evidence. Evaluation of 30 general‑purpose and medical LLMs shows that strong diagnostic accuracy does not translate into better downstream care, with failure modes shifting from explicit decision errors in staging to decision omissions in initial work‑up and follow‑up, and reveals a trade‑off between prompt granularity, decision quality, evidence faithfulness, and safety adherence.