Recent breakthroughs in generative AI have opened the door to new research perspectives in the domain of art and cultural heritage, where a large number of artifacts have been digitized. There is a need for innovation to ease the access and highlight the content of digital collections. Such innovations develop into creative explorations of the digital image in relation to its malleability and contemporary interpretation, in confrontation to the original historical object. Based on the concept of the autonomous image, we propose a new framework towards the production of self-explaining cultural artifacts using open-source large-language, face detection, text-to-speech and audio-to-animation models. The goal is to start from a digitized artwork and to automatically assemble a short video of the latter where the main character animates to explain its content. The whole process questions cultural biases encapsulated in large-language models, the potential of digital images and deepfakes of artworks for educational purposes, along with concerns of the field of art history regarding such creative diversions.

Multimodal Large Language Models (MLLMs) like GPT-4V are capable of reasoning across text and image modalities, showing promise in a variety of complex vision-language tasks. In this preliminary study, we investigate the out-of-the-box capabilities of GPT-4V in the domain of image forensics, specifically, in detecting image splicing manipulations. Without any task-specific fine-tuning, we evaluate GPT-4V using three prompting strategies: Zero-Shot (ZS), Few-Shot (FS), and Chain-of-Thought (CoT), applied over a curated subset of the CASIA v2.0 splicing dataset. Our results show that GPT-4V achieves competitive detection performance in zero-shot settings (more than 85% accuracy), with CoT prompting yielding the most balanced trade-off across authentic and spliced images. Qualitative analysis further reveals that the model not only detects low-level visual artifacts but also draws upon real-world contextual knowledge such as object scale, semantic consistency, and architectural facts, to identify implausible composites. While GPT-4V lags behind specialized state-of-the-art splicing detection models, its generalizability, interpretability, and encyclopedic reasoning highlight its potential as a flexible tool in image forensics.

Large reasoning models (LRMs) achieve higher performance on challenging reasoning tasks by generating more tokens at inference time, but this verbosity often wastes computation on easy problems. Existing solutions, including supervised finetuning on shorter traces, user-controlled budgets, or RL with uniform penalties, either require data curation, manual configuration, or treat all problems alike regardless of difficulty. We introduce Adaptive Length Penalty (ALP), a reinforcement learning objective tailoring generation length to per-prompt solve rate. During training, ALP monitors each prompt's online solve rate through multiple rollouts and adds a differentiable penalty whose magnitude scales inversely with that rate, so confident (easy) prompts incur a high cost for extra tokens while hard prompts remain unhindered. Posttraining DeepScaleR-1.5B with ALP cuts average token usage by 50\% without significantly dropping performance. Relative to fixed-budget and uniform penalty baselines, ALP redistributes its reduced budget more intelligently by cutting compute on easy prompts and reallocating saved tokens to difficult ones, delivering higher accuracy on the hardest problems with higher cost.

Large Language Models (LLMs) have shown remarkable performance in Open-Domain Question Answering (ODQA) by leveraging external documents through Retrieval-Augmented Generation (RAG). To reduce RAG overhead, from longer context, context compression is necessary. However, prior compression methods do not focus on filtering out non-evidential information, which limit the performance in LLM-based RAG. We thus propose Evidentiality-guided RAG, or ECoRAG framework. ECoRAG improves LLM performance by compressing retrieved documents based on evidentiality, ensuring whether answer generation is supported by the correct evidence. As an additional step, ECoRAG reflects whether the compressed content provides sufficient evidence, and if not, retrieves more until sufficient. Experiments show that ECoRAG improves LLM performance on ODQA tasks, outperforming existing compression methods. Furthermore, ECoRAG is highly cost-efficient, as it not only reduces latency but also minimizes token usage by retaining only the necessary information to generate the correct answer. Code is available at https://github.com/ldilab/ECoRAG.

Large Language Models (LLMs) are known to exhibit social, demographic, and gender biases, often as a consequence of the data on which they are trained. In this work, we adopt a mechanistic interpretability approach to analyze how such biases are structurally represented within models such as GPT-2 and Llama2. Focusing on demographic and gender biases, we explore different metrics to identify the internal edges responsible for biased behavior. We then assess the stability, localization, and generalizability of these components across dataset and linguistic variations. Through systematic ablations, we demonstrate that bias-related computations are highly localized, often concentrated in a small subset of layers. Moreover, the identified components change across fine-tuning settings, including those unrelated to bias. Finally, we show that removing these components not only reduces biased outputs but also affects other NLP tasks, such as named entity recognition and linguistic acceptability judgment because of the sharing of important components with these tasks.

Large language models (LLMs) exhibit compelling linguistic behaviour, and sometimes offer self-reports, that is to say statements about their own nature, inner workings, or behaviour. In humans, such reports are often attributed to a faculty of introspection and are typically linked to consciousness. This raises the question of how to interpret self-reports produced by LLMs, given their increasing linguistic fluency and cognitive capabilities. To what extent (if any) can the concept of introspection be meaningfully applied to LLMs? Here, we present and critique two examples of apparent introspective self-report from LLMs. In the first example, an LLM attempts to describe the process behind its own "creative" writing, and we argue this is not a valid example of introspection. In the second example, an LLM correctly infers the value of its own temperature parameter, and we argue that this can be legitimately considered a minimal example of introspection, albeit one that is (presumably) not accompanied by conscious experience.

We consider the Top-$K$ selection problem, which aims to identify the largest-$K$ elements from an array. Top-$K$ selection arises in many machine learning algorithms and often becomes a bottleneck on accelerators, which are optimized for dense matrix multiplications. To address this problem, \citet{chern2022tpuknnknearestneighbor} proposed a fast two-stage \textit{approximate} Top-$K$ algorithm: (i) partition the input array and select the top-$1$ element from each partition, (ii) sort this \textit{smaller subset} and return the top $K$ elements. In this paper, we consider a generalized version of this algorithm, where the first stage selects top-$K'$ elements, for some $1 \leq K' \leq K$, from each partition. Our contributions are as follows: (i) we derive an expression for the expected recall of this generalized algorithm and show that choosing $K' > 1$ with fewer partitions in the first stage reduces the input size to the second stage more effectively while maintaining the same expected recall as the original algorithm, (ii) we derive a bound on the expected recall for the original algorithm in \citet{chern2022tpuknnknearestneighbor} that is provably tighter by a factor of $2$ than the one in that paper, and (iii) we implement our algorithm on Cloud TPUv5e and achieve around an order of magnitude speedups over the original algorithm without sacrificing recall on real-world tasks.

Factuality is a necessary precursor to useful educational tools. As adoption of Large Language Models (LLMs) in education continues of grow, ensuring correctness in all settings is paramount. Despite their strong English capabilities, LLM performance in other languages is largely untested. In this work, we evaluate the correctness of the Llama3.1 family of models in answering factual questions appropriate for middle and high school students. We demonstrate that LLMs not only provide extraneous and less truthful information, but also exacerbate existing biases against rare languages.

With the recent advances in Artificial Intelligence (AI) and Large Language Models (LLMs), the automation of daily tasks, like automatic writing, is getting more and more attention. Hence, efforts have focused on aligning LLMs with human values, yet humor, particularly professional industrial humor used in workplaces, has been largely neglected. To address this, we develop a dataset of professional humor statements along with features that determine the appropriateness of each statement. Our evaluation of five LLMs shows that LLMs often struggle to judge the appropriateness of humor accurately.

Detoxification, the task of rewriting harmful language into non-toxic text, has become increasingly important amid the growing prevalence of toxic content online. However, high-quality parallel datasets for detoxification, especially for hate speech, remain scarce due to the cost and sensitivity of human annotation. In this paper, we propose a novel LLM-in-the-loop pipeline leveraging GPT-4o-mini for automated detoxification. We first replicate the ParaDetox pipeline by replacing human annotators with an LLM and show that the LLM performs comparably to human annotation. Building on this, we construct ParaDeHate, a large-scale parallel dataset specifically for hatespeech detoxification. We release ParaDeHate as a benchmark of over 8K hate/non-hate text pairs and evaluate a wide range of baseline methods. Experimental results show that models such as BART, fine-tuned on ParaDeHate, achieve better performance in style accuracy, content preservation, and fluency, demonstrating the effectiveness of LLM-generated detoxification text as a scalable alternative to human annotation.