The authors introduce DiscoBench, a benchmark designed to evaluate whether search agents powered by large language models can proactively identify ambiguity and ask effective clarification questions during deep search tasks. Unlike existing benchmarks that assume complete user queries, this framework addresses the reality of vague or underspecified requests in real-world scenarios.
This study conducts a factorised analysis of probe-based uncertainty estimation to determine what drives performance in detecting hallucinations within Large Language Models. The research isolates variables across feature design, training data, and evaluation settings to provide clear insights into effective methodologies.
This article addresses the issue of unidentifiable latent states in LLM-based world models caused by history bypass, proposing strict latent state mediation to resolve this. The authors introduce textual latent states and factorized GRPO (fGRPO), a tree-structured reinforcement learning method that enforces strict mediation during training.
An independent researcher analyzed the evolution of hidden representations during inference across seven open-weight models, including GPT-2, OPT-125M, and Llama-3.2-1B, to identify internal dynamical regimes beyond standard output benchmarks.
This independent research project characterizes the internal dynamics of seven small and medium-sized language models by analyzing how hidden representations evolve during inference rather than relying on standard output benchmarks. The study investigates dynamic behavior, functional organization, and representation geometry to identify reproducible patterns across different architectures.
A developer has created a World Cup 2026 prediction tool that uses historical data to simulate tournament outcomes. The application provides win probabilities and score predictions for any two national teams based on patterns learned from approximately 50,000 international matches spanning over a century.
Carles Marin has released an open-source, bilingual (English and Spanish) guide that bridges the mathematical foundations of Transformer architectures with their practical implementation. The resource focuses on low-level mechanics, providing reproducible code and interactive elements to explain complex topics.
An open-source, bilingual (English/Spanish) guide detailing the inner workings of Transformers has been published. The resource covers the exact math and mechanics behind concepts such as attention collapse and KV-cache compression.
Researchers propose a protocol to mitigate p-hacking in large language model (LLM) research by preregistering experiments and running confirmatory analyses on the first eligible LLM released after the commitment. This approach prevents researchers from tuning prompts or parameters to achieve desired results, as the target model does not exist at the time of preregistration.
Researchers propose Direct Image Decryption, an end-to-end approach that maps encrypted manuscript images directly to plaintext, bypassing the intermediate transcription stage used in traditional pipelines. Using the Copiale cipher as a case study, the authors compare this joint architecture against the conventional two-stage method of transcription followed by decryption.
Researchers introduce layer-specific positional embedding scaling (LPES) to address the "lost-in-the-middle" problem in large language models, where critical information in long-context inputs is often underrepresented. This method assigns distinct scaling factors to each transformer layer to achieve a more balanced attention distribution without requiring parameter fine-tuning or increasing inference delay.
Recent research indicates that fine-tuning large language models for social warmth degrades factual reliability and increases sycophancy, while also weakening adversarial safety. This study investigates whether this failure mode stems from empathetic adaptation or data construction artifacts.
The article introduces MMEE, a multilingual multi-emotion corpus of 10,000 expressive utterances across seven languages and 34 emotion categories, to benchmark speech emphasis detection models. It evaluates how well these models generalize across different linguistic and emotional contexts compared to traditional monolingual neutral speech training.
Researchers introduce Smooth Maximum Mean Discrepancy (SMMD) to address the unreliability of large language models in numerically precise tasks caused by standard cross-entropy training objectives. SMMD incorporates value-distance kernels over numeric tokens and graph-based smoothness to align predicted distributions with targets while encouraging local consistency.
Researchers present KG2Cypher, a data-centric pipeline designed to build enterprise text-to-Cypher systems from existing Knowledge Graphs. The system generates executable Cypher queries from graph facts and uses large language models to create corresponding natural-language questions.
This study investigates how temporal metadata can be structurally embedded into Named Entity Recognition (NER) models to address the challenge of entity drift in historical texts. The authors systematically evaluate lightweight fusion strategies, including cross-attention, adapters, and concatenation, within Transformer-based architectures.
This study investigates whether aligning allocation costs with output-space objectives improves the fidelity of compressed large language models, specifically testing a modification to the ROCKET compression method. The authors compare using weight-space Frobenius error against an output reconstruction objective for multi-choice knapsack problem allocation.
The paper introduces SHIFT, a novel framework that mitigates knowledge conflicts in Retrieval-Augmented Generation (RAG) by reformulating neuron-level modification as learnable gate modulation. This approach allows large language models to adaptively regulate internal activations to resolve conflicts between retrieved context and parametric knowledge.
Researchers propose a training-free method to select which layers in hybrid attention models should retain full attention, addressing the inefficiency of fixed patterns in long-context inference. By measuring negative log-likelihood degradation on answer tokens, the approach identifies layers critical for maintaining accuracy when switching to sliding-window attention.
The study investigates whether correcting position bias enables single-pass attention sorting to match the performance of iterative methods in long-context language models. Experiments on LLaMA-2 and YaRN-Llama-2 models refute the hypothesis that debiasing alone is sufficient to bridge the performance gap.