Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
Defeasible reasoning is a type of reasoning where inferences are drawn from plausible current evidence, but can be retracted upon the introduction of newer evidence. Although recent studies have exami...
Multi-Agent Debate (MAD) has been widely adopted to improve LLM-based evaluation by prompting multiple agents to negotiate and reach a consensus. However, for subjective rubric-based scoring, inter-ag...
OpenAI supports California SB 1119, advancing strong, age-appropriate AI safeguards for teens while preserving opportunities to learn, create, and explore.
General embodied agents should perceive, predict, act, evaluate, and improve within a unified system. World models have shown great promise in building such agents, yet existing models typically appen...
Automatic medical coding assigns ICD codes to clinical notes, but it remains challenging due to long documents, imbalanced label distributions, and diverse terms. These challenges are especially sever...
We propose a new design of fair classifiers for multi-class classification problems in the presence of vector-valued sensitive attributes. In that scenario each sensitive attribute has multiple values...
SymboLLM-FE combines symbolic regression and large language models to automate feature engineering for tabular data. It first extracts mathematically expressive formulas that correlate strongly with the target, then refines them with LLMs to improve interpretability. Experiments on six real‑world datasets and four Kaggle competitions show that SymboLLM‑FE outperforms existing AutoFE methods while reducing the number of costly LLM calls.
Fast Weight Attention for Continual Learning introduces recurrent fast‑weight memories and selective state‑space models that compress expanding context into a fixed‑size recurrent state, enabling an online learning rule for state transitions. The paper derives normalized first‑order updates for squared‑error regression and negative inner‑product objectives, presenting several variants (Falcon‑1, Falcon‑2, Falcon‑3 and their inner‑product counterparts) with recurrent, masked‑parallel, and chunk‑parallel implementations. These methods demonstrate competitive performance in language modeling and improved length extrapolation on variable‑digit addition tasks.
The paper introduces a depth‑aware pothole detection framework that fuses RGB‑D sensor data and evaluates five architectures—YOLOv8n, YOLOv8nSeg, YOLOv9t, RTDETRL, and RTDETRX—on the PothRGBD dataset. YOLOv8nSeg achieves the highest detection performance (mAP@50 = 0.9556, mAP@50_95 = 0.6758) and the most accurate depth estimate (2.96 cm), while YOLOv8n offers the fastest inference (3.6 ms) and RTDETRX delivers the highest detection confidence (92.70 %). The study also shows that even after RANSAC orthorectification, bounding‑box models overestimate pothole depth by 0.16–0.21 cm, indicating a structural bias rather than a calibration error.
The study evaluates machine learning models for nitrogen recommendations in winter wheat by directly scoring profit loss on 892 yield response curves, rather than relying on prediction accuracy. Results show that none of the models recover the best rate within farm tolerance, and at typical prices all models underperform standard UK advice. A simple post‑model correction step reduces profit losses by up to 43% without retraining, while a hybrid approach further mitigates bias and large losses.
The paper introduces the Question Damage Score, a diagnostic framework that evaluates how much large language models rely on specific context examples in linguistic reasoning tasks. By systematically deleting either random or strategically chosen context examples from 53 UK Linguistics Olympiad puzzles, the authors classify puzzles as fragile or robust and assess model performance under instructions to abstain when information is insufficient. The study finds that current frontier LLMs rarely abstain and often still answer correctly even after load‑bearing context is removed, highlighting a need to better understand context‑based reasoning, prior knowledge, and memorization in these models.
The paper introduces the Persuasion Index (PI), a taxonomy of 15 persuasion dimensions grounded in psychological and communication theories, implemented with 55 lexicon- and rule-based sub-features. PI is modular, allowing individual detectors to be swapped while preserving its theoretical framework. Evaluations on four English argumentative datasets show that PI provides a shared, lightweight feature space that captures meaningful predictive signals and reveals consistent dimension-level associations with persuasion outcomes, with variations across topics and stances.
The paper investigates how different proper scoring rules influence the performance and behavior of large language model (LLM) forecasters. Five scoring rules were compared as training objectives for binary forecasts of real-world events, revealing that while they all theoretically incentivize truthful probability reporting, they produce models with varying calibration, probability usage, and bias, information, and noise profiles. The Brier-trained model achieved the lowest Brier score and highest AUC-ROC, whereas the log-trained model achieved the best log score and lowest calibration error, indicating that scoring rule choice can shape both forecast accuracy and error structure.
The article surveys AI alignment from a game-theoretic perspective, focusing on how large language models and AI agents can be aligned with complex human values in high-risk settings. It categorizes recent progress around key game-theoretic elements and addresses three main challenges: preference diversity, alignment priority, and temporal dynamics. The survey clarifies where game theory benefits current alignment methods, where its application is looser, and what remains to be tackled for robust, adaptive, and verifiable AI systems.
The paper introduces AIM, a two‑stage approach for unlearning identity‑specific information from multimodal large language models (MLLMs) when retain images are not available at deletion time. AIM first anchors an identity‑forgetting target using a universal visual prompt, then aligns the vision encoder to this target under a Fisher‑based constraint. Experiments demonstrate that AIM effectively removes identity knowledge while preserving other visual perception capabilities and prior knowledge.
The paper shows that unsupervised representational alignment can fail due to symmetry in the stimulus geometry, even before data are collected. By using a design-time diagnostic based on the automorphism group of the geometry, the authors demonstrate that dense sampling can create near-duplicates that make alignment degenerate. Applying this diagnostic to a colour design reduces catastrophic alignment failures from 75% to 2% without altering models, layers, or solvers.
The paper investigates affective framing in Bengali digital journalism by analyzing 300,000 news headlines with zero‑shot inference using Gemma 3 4B to estimate dominant emotions and overall affective tone. Results reveal frequent negative emotions—especially anger, sadness, disappointment, and fear—across the corpus, and a pilot validation on 200 headlines indicates the model’s estimates are useful yet not definitive benchmarks. The authors suggest a bias‑sensitive news interface that visualizes emotional cues to help readers detect affective framing patterns in daily news.
The paper introduces DARTS, a method for tuning decoder representations during model merging. It addresses representation bias in autoregressive decoders by using an entropy‑weighted L1 loss and a per‑position additive bias to correct errors that accumulate across token positions. Experiments on code generation, mathematical reasoning, and instruction following with Llama‑2‑7B show that DARTS improves performance over standard surgery while adding only 0.1% extra parameters.
The paper introduces a privacy‑preserving zk‑SNARK audit framework that uses adversarial‑style probes to detect logit drift between an approved large language model and a modified deployment. It offers three probe families—token‑based (black‑box), embedding‑based (gray‑box), and stress probes (partial white‑box)—allowing users to balance sensitivity, access, and cost. Experiments across LLM architectures and GPU platforms show token‑based probes achieve the highest mean sensitivity while remaining practical in a black‑box setting, with Groth16 proving times scaling modestly from 1.02 to 1.78 seconds and constant proof size.
The paper introduces EXPO, an algorithm that automatically optimizes the meta-prompt—specifically the task description and meta-instruction—for large language model agents in sequential decision-making tasks such as Bayesian optimization and multi-armed bandits. Building on adversarial bandit techniques to handle non-stationary rewards, the authors extend EXPO to EXPO-ES, which also optimizes exemplars (historical interactions) within the meta-prompt. Experiments demonstrate that these methods significantly improve the performance of LLM-based agents in sequential decision-making scenarios.