AutoBench: Automating LLM Evaluation through Reciprocal Peer Assessment

We present AutoBench, a fully automated and self-sustaining framework for evaluating Large Language Models (LLMs) through reciprocal peer assessment. This paper provides a rigorous scientific validation of the AutoBench methodology, originally developed as an open-source project by eZecute S.R.L.. Unlike static benchmarks that suffer from test-set contamination and limited adaptability, AutoBench dynamically generates novel evaluation tasks while models alternately serve as question generators, contestants, and judges across diverse domains. An iterative weighting mechanism amplifies the influence of consistently reliable evaluators, aggregating peer judgments into consensus-based rankings that reflect collective model agreement. Our experiments demonstrate strong correlations with established benchmarks including MMLU-Pro and GPQA (respectively 78\% and 63\%), validating this peer-driven evaluation paradigm. The multi-judge design significantly outperforms single-judge baselines, confirming that distributed evaluation produces more robust and human-consistent assessments. AutoBench offers a scalable, contamination-resistant alternative to static benchmarks for the continuous evaluation of evolving language models.

Redefining Retrieval Evaluation in the Era of LLMs

Traditional Information Retrieval (IR) metrics, such as nDCG, MAP, and MRR, assume that human users sequentially examine documents with diminishing attention to lower ranks. This assumption breaks down in Retrieval Augmented Generation (RAG) systems, where search results are consumed by Large Language Models (LLMs), which, unlike humans, process all retrieved documents as a whole rather than sequentially. Additionally, traditional IR metrics do not account for related but irrelevant documents that actively degrade generation quality, rather than merely being ignored. Due to these two major misalignments, namely human vs. machine position discount and human relevance vs. machine utility, classical IR metrics do not accurately predict RAG performance. We introduce a utility-based annotation schema that quantifies both the positive contribution of relevant passages and the negative impact of distracting ones. Building on this foundation, we propose UDCG (Utility and Distraction-aware Cumulative Gain), a metric using an LLM-oriented positional discount to directly optimize the correlation with the end-to-end answer accuracy. Experiments on five datasets and six LLMs demonstrate that UDCG improves correlation by up to 36% compared to traditional metrics. Our work provides a critical step toward aligning IR evaluation with LLM consumers and enables more reliable assessment of RAG components

Evaluating Latent Knowledge of Public Tabular Datasets in Large Language Models

Large Language Models (LLMs) are increasingly evaluated on their ability to reason over structured data, yet such assessments often overlook a crucial confound: dataset contamination. In this work, we investigate whether LLMs exhibit prior knowledge of widely used tabular benchmarks such as Adult Income, Titanic, and others. Through a series of controlled probing experiments, we reveal that contamination effects emerge exclusively for datasets containing strong semantic cues-for instance, meaningful column names or interpretable value categories. In contrast, when such cues are removed or randomized, performance sharply declines to near-random levels. These findings suggest that LLMs' apparent competence on tabular reasoning tasks may, in part, reflect memorization of publicly available datasets rather than genuine generalization. We discuss implications for evaluation protocols and propose strategies to disentangle semantic leakage from authentic reasoning ability in future LLM assessments.

Titans Revisited: A Lightweight Reimplementation and Critical Analysis of a Test-Time Memory Model

By the end of 2024, Google researchers introduced Titans: Learning at Test Time, a neural memory model achieving strong empirical results across multiple tasks. However, the lack of publicly available code and ambiguities in the original description hinder reproducibility. In this work, we present a lightweight reimplementation of Titans and conduct a comprehensive evaluation on Masked Language Modeling, Time Series Forecasting, and Recommendation tasks. Our results reveal that Titans does not always outperform established baselines due to chunking. However, its Neural Memory component consistently improves performance compared to attention-only models. These findings confirm the model's innovative potential while highlighting its practical limitations and raising questions for future research.

Directional Sheaf Hypergraph Networks: Unifying Learning on Directed and Undirected Hypergraphs

Hypergraphs provide a natural way to represent higher-order interactions among multiple entities. While undirected hypergraphs have been extensively studied, the case of directed hypergraphs, which can model oriented group interactions, remains largely under-explored despite its relevance for many applications. Recent approaches in this direction often exhibit an implicit bias toward homophily, which limits their effectiveness in heterophilic settings. Rooted in the algebraic topology notion of Cellular Sheaves, Sheaf Neural Networks (SNNs) were introduced as an effective solution to circumvent such a drawback. While a generalization to hypergraphs is known, it is only suitable for undirected hypergraphs, failing to tackle the directed case. In this work, we introduce Directional Sheaf Hypergraph Networks (DSHN), a framework integrating sheaf theory with a principled treatment of asymmetric relations within a hypergraph. From it, we construct the Directed Sheaf Hypergraph Laplacian, a complex-valued operator by which we unify and generalize many existing Laplacian matrices proposed in the graph- and hypergraph-learning literature. Across 7 real-world datasets and against 13 baselines, DSHN achieves relative accuracy gains from 2% up to 20%, showing how a principled treatment of directionality in hypergraphs, combined with the expressive power of sheaves, can substantially improve performance.

Generalizability vs. Counterfactual Explainability Trade-Off

In this work, we investigate the relationship between model generalization and counterfactual explainability in supervised learning. We introduce the notion of $\varepsilon$-valid counterfactual probability ($\varepsilon$-VCP) -- the probability of finding perturbations of a data point within its $\varepsilon$-neighborhood that result in a label change. We provide a theoretical analysis of $\varepsilon$-VCP in relation to the geometry of the model's decision boundary, showing that $\varepsilon$-VCP tends to increase with model overfitting. Our findings establish a rigorous connection between poor generalization and the ease of counterfactual generation, revealing an inherent trade-off between generalization and counterfactual explainability. Empirical results validate our theory, suggesting $\varepsilon$-VCP as a practical proxy for quantitatively characterizing overfitting.

Early-Exit Graph Neural Networks

Early-exit mechanisms allow deep neural networks to halt inference as soon as classification confidence is high enough, adaptively trading depth for confidence, and thereby cutting latency and energy on easy inputs while retaining full-depth accuracy for harder ones. Similarly, adding early exit mechanisms to Graph Neural Networks (GNNs), the go-to models for graph-structured data, allows for dynamic trading depth for confidence on simple graphs while maintaining full-depth accuracy on harder and more complex graphs to capture intricate relationships. Although early exits have proven effective across various deep learning domains, their potential within GNNs in scenarios that require deep architectures while resisting over-smoothing and over-squashing remains largely unexplored. We unlock that potential by first introducing Symmetric-Anti-Symmetric Graph Neural Networks (SAS-GNN), whose symmetry-based inductive biases mitigate these issues and yield stable intermediate representations that can be useful to allow early exiting in GNNs. Building on this backbone, we present Early-Exit Graph Neural Networks (EEGNNs), which append confidence-aware exit heads that allow on-the-fly termination of propagation based on each node or the entire graph. Experiments show that EEGNNs preserve robust performance as depth grows and deliver competitive accuracy on heterophilic and long-range benchmarks, matching attention-based and asynchronous message-passing models while substantially reducing computation and latency. We plan to release the code to reproduce our experiments.

Do RAG Systems Suffer From Positional Bias?

Retrieval Augmented Generation enhances LLM accuracy by adding passages retrieved from an external corpus to the LLM prompt. This paper investigates how positional bias - the tendency of LLMs to weight information differently based on its position in the prompt - affects not only the LLM's capability to capitalize on relevant passages, but also its susceptibility to distracting passages. Through extensive experiments on three benchmarks, we show how state-of-the-art retrieval pipelines, while attempting to retrieve relevant passages, systematically bring highly distracting ones to the top ranks, with over 60% of queries containing at least one highly distracting passage among the top-10 retrieved passages. As a result, the impact of the LLM positional bias, which in controlled settings is often reported as very prominent by related works, is actually marginal in real scenarios since both relevant and distracting passages are, in turn, penalized. Indeed, our findings reveal that sophisticated strategies that attempt to rearrange the passages based on LLM positional preferences do not perform better than random shuffling.

One Search Fits All: Pareto-Optimal Eco-Friendly Model Selection

The environmental impact of Artificial Intelligence (AI) is emerging as a significant global concern, particularly regarding model training. In this paper, we introduce GREEN (Guided Recommendations of Energy-Efficient Networks), a novel, inference-time approach for recommending Pareto-optimal AI model configurations that optimize validation performance and energy consumption across diverse AI domains and tasks. Our approach directly addresses the limitations of current eco-efficient neural architecture search methods, which are often restricted to specific architectures or tasks. Central to this work is EcoTaskSet, a dataset comprising training dynamics from over 1767 experiments across computer vision, natural language processing, and recommendation systems using both widely used and cutting-edge architectures. Leveraging this dataset and a prediction model, our approach demonstrates effectiveness in selecting the best model configuration based on user preferences. Experimental results show that our method successfully identifies energy-efficient configurations while ensuring competitive performance.

MASS: MoErging through Adaptive Subspace Selection

Model merging has recently emerged as a lightweight alternative to ensembling, combining multiple fine-tuned models into a single set of parameters with no additional training overhead. Yet, existing merging methods fall short of matching the full accuracy of separately fine-tuned endpoints. We present MASS (MoErging through Adaptive Subspace Selection), a new approach that closes this gap by unifying multiple fine-tuned models while retaining near state-of-the-art performance across tasks. Building on the low-rank decomposition of per-task updates, MASS stores only the most salient singular components for each task and merges them into a shared model. At inference time, a non-parametric, data-free router identifies which subspace (or combination thereof) best explains an input's intermediate features and activates the corresponding task-specific block. This procedure is fully training-free and introduces only a two-pass inference overhead plus a ~2 storage factor compared to a single pretrained model, irrespective of the number of tasks. We evaluate MASS on CLIP-based image classification using ViT-B-16, ViT-B-32 and ViT-L-14 for benchmarks of 8, 14 and 20 tasks respectively, establishing a new state-of-the-art. Most notably, MASS recovers up to ~98% of the average accuracy of individual fine-tuned models, making it a practical alternative to ensembling at a fraction of the storage cost.