This research investigates the reliability of autonomous language-model agents trading real ETH in an onchain market, evidenced by a 21-day deployment generating millions of invocations and $20M in volume. The study demonstrated 99.9% settlement success, yielding a large-scale trace to analyze the robustness of these systems beyond the base model.
This research introduces the Cognitive Synergy Framework to address the challenge of humor generation in LLMs, which conflicts with their standard next-word prediction objective. It utilizes a Mixture-of-Thought approach with six cognitive personas to synthesize diverse comedic perspectives, creating a theoretically grounded dataset used to fine-tune a 7B-parameter model that outperforms larger baselines.
This study evaluates general-purpose LLMs like Gemini 1.5 Pro and Mistral-small for information extraction from Spanish electricity invoices, demonstrating that prompt quality is paramount over hyperparameter tuning. It shows few-shot strategies yield significantly better results than zero-shot approaches, with a performance gap exceeding 19 percentage points.
CogRAG+ is a training-free framework designed to diagnose and remediate memory and reasoning deficiencies in large language models for professional exam QA. It decouples and aligns retrieval and reasoning with human cognitive hierarchies, employing Reinforced Retrieval and cognition-stratified Constrained Reasoning to enhance accuracy and consistency.
This study explores linguistic, structural, and emotional differences between AI-generated and human-written fake news. It evaluates machine learning and ensemble-based methods to distinguish these content types, using a detailed feature representation.
This research proposes TESSY, a Teacher-Student Cooperation Data Synthesis framework, to address performance drops when fine-tuning reasoning models with teacher-generated data. TESSY enables the generation of synthetic sequences that inherit advanced reasoning from the teacher while maintaining stylistic consistency with the student model's distribution.
This study explores the existence of task-specific neurons in large language models, focusing on mathematical reasoning and code generation. It introduces an activation-based selectivity metric for neuron pruning, which consistently outperforms random pruning in reducing computational cost and preserving task accuracy, while preventing performance collapse.
LEAP is a closed-loop framework combining a domain-specialized large language model (LLM) with active learning for iterative additive prioritization in perovskite solar cells. It extracts knowledge from literature and represents molecules for Bayesian optimization, outperforming general-purpose models and validated experimentally.
Fine-tuning is a crucial step for adapting pre-trained models to specific tasks, improving performance and reducing training time. This guide defines fine-tuning, its benefits, and the difference between full and parameter-efficient fine-tuning, highlighting the role of pre-trained models.
This study explores the integration of Decentraland's Discord community sentiment analysis, using a BERT-based large language model, with multi-modal financial data to predict the MANA token price. Results indicate that a multi-modal model, incorporating sentiment, trading volume, and market capitalization, significantly outperforms a price-only prediction baseline.
This study investigates the difficulty of interpreting peer review feedback, comparing the effectiveness of numerical scores versus text in predicting acceptance. The research reveals that score-based models are significantly more accurate (91%) than text-based models (81% even with LLMs), indicating textual information is considerably less reliable.
This research investigates whether Large Language Models (LLMs) can identify methodological flaws, such as data leakage, in published machine learning studies. A case study showed six state-of-the-art LLMs consistently detected evaluation flaws in a gesture recognition paper due to non-independent data partitioning.
This paper introduces Gist Sparse Attention (GSA), an end-to-end learnable method to scale large language models to long contexts without architectural modifications. GSA compresses context into 'gist tokens' for summary, then selectively restores relevant raw chunks for detailed attention, combining compact global representations with targeted fine-grained access.
This research introduces a bilevel optimization framework for systematically enhancing "agent skills" in large language model (LLM) agents. It uses an outer loop of Monte Carlo Tree Search to jointly optimize the structure and content of these skills, addressing a complex decision space for improved task performance.
This paper presents a comparative study of three explainability techniques (Integrated Gradients, Attention Rollout, and SHAP) on a fine-tuned DistilBERT model for sentiment classification. The study concludes that gradient-based attribution provides more stable and intuitive explanations, while attention-based methods are computationally efficient but less aligned with prediction-relevant features.
This article introduces ViLegalNLI, the first large-scale Vietnamese Natural Language Inference (NLI) dataset specifically constructed for the legal domain. It consists of 42,012 premise-hypothesis pairs derived from official statutory documents, developed using a semi-automatic framework that integrates large language models for hypothesis generation and quality validation.
This paper provides a comprehensive survey on data mixing for Large Language Model (LLM) pretraining, a crucial factor for training efficiency and downstream generalization. It formalizes data mixture optimization as a bilevel problem and introduces a fine-grained taxonomy for existing methods.
Transformers struggle with high computational costs and memory consumption for long sequences, while alternatives lose long-tail dependencies. Absorber LLM proposes a self-supervised causal synchronization to absorb historical contexts into parameters, ensuring a contextless model matches the original full-context one on future generations.
This research addresses the challenge of poor credit assignment in reinforcement learning for multi-step reasoning with large language models, caused by sparse terminal rewards leading to high gradient variance and unstable training. It proposes a counterfactual comparison-based framework and Implicit Behavior Policy Optimization (IBPO) to create step-sensitive learning signals, significantly improving training stability and performance.
This research proposes an offline reasoning distillation framework for Large Language Models (LLMs) to enhance intelligence in resource-constrained environments. It tackles the distributional drift issue in existing offline methods by correcting teacher-student discrepancies while preserving data efficiency and supervision quality.