This research introduces EdgeRazor, a lightweight framework designed to deploy Large Language Models on resource-constrained devices. It leverages mixed-precision quantization-aware distillation to convert full-precision models into lower-bit formats, overcoming limitations of previous quantization methods.
MemQ integrates TD($\lambda$) eligibility traces with memory Q-values, propagating credit backward through a provenance DAG to account for memory dependencies. This approach significantly improves LLM agents' ability to accumulate and retrieve experience, achieving high success rates across various benchmarks.
The paper introduces SMDD-Bench, a new challenging multi-turn benchmark with 502 guaranteed-solvable tasks to evaluate LLM agents' performance in real-world small molecule drug design. It aims to standardize evaluation across diverse chemistries and targets, requiring strong chemical, biological, and 3D intuition.
This research proposes distinguishing between capability elicitation and capability creation in large language model post-training. It argues that elicitation reweights existing behaviors within a model's accessible support, while creation changes that support itself, developing this through a free-energy view.
This research explores whether LLMs can serve as a lens for understanding neural representations of emotional valence in the human brain, focusing on EEG. It builds a valence axis from LLMs and demonstrates its mapping onto human neural activity, suggesting a shared representation.
This paper introduces Deployment-Time Learning (DTL) as a new stage for LLMs, allowing them to continually adapt from experience post-training without modifying core parameters. It presents CASCADE, a framework that uses an explicit, evolving episodic memory for LLM agents, formalizing experience reuse as a contextual bandit problem with no-regret guarantees.
This research introduces MOOD, a benchmark designed to study the detection of out-of-distribution (OOD) alignment failures in large language models (LLMs) using monitoring pipelines. It proposes combining guard models with OOD detectors to improve the generalization of safety classifiers, which often fail in OOD scenarios.
This paper introduces GraphDC, a Divide-and-Conquer multi-agent system designed to enhance graph algorithm reasoning in Large Language Models (LLMs). It improves performance by decomposing large graphs into smaller subgraphs for specialized agents, with a master agent integrating the results, leading to better scalability and robustness.
This research paper introduces a coin-flip model to predict the performance of symbolic and prompt-based LLM programs using a few in-domain examples and a performance prior. It finds that symbolic programs exhibit an "all or nothing" performance prior, while prompt programs have a diffuse prior.
This paper investigates how LLMs learn in-context, using a graph random-walk task to explore whether they pattern-match or infer latent structure. It reveals that neither account alone is sufficient, presenting evidence of simultaneous encoding of graph topologies and causal interventions.
This study introduces AgentNLQ, a new multi-agent method for Natural Language to SQL (NL2SQL) conversion, achieving 78.1% semantic accuracy on the BIRD benchmark. It leverages LLMs in an optimized orchestrator for planning, reflection, and self-correction to generate accurate SQL queries from enriched schemas and business rules.
This paper introduces a new paradigm for interactively evaluating Theory of Mind (ToM) improvements in Large Language Models (LLMs) for human-AI interactions. Empirical findings from real-world datasets and a user study reveal that ToM enhancements on static benchmarks do not always translate to benefits in dynamic human-AI interactions.
This research evaluates large language models (LLMs) in biomedical question answering, specifically addressing their reliability when faced with conflicting or incomplete evidence. It reveals that LLM accuracy significantly drops, and predictions flip, when the order of correct and contradictory documents is reversed, highlighting issues with order effects and the need for conflict-aware abstention.
This research paper presents an atlas of domain-level metacognitive monitoring across 33 frontier LLMs, analyzing 1,500 MMLU items across six domains. It reveals significant within-model variation, with Applied/Professional knowledge being the easiest and Formal Reasoning/Natural Science the hardest domains to monitor.
This paper introduces CAX-Agent, a lightweight agent harness designed to enhance the reliability of large language models (LLMs) in MAPDL finite-element simulations. It addresses issues like inconsistent outputs and task failures through structured execution control, tool encapsulation, and robust fault recovery mechanisms, evaluating various recovery strategies.
This paper argues that current Uncertainty Quantification (UQ) methods for LLMs are essentially unsupervised clustering algorithms, measuring internal consistency rather than external correctness. Consequently, these methods fail to detect "confident hallucinations" and may create a deceptive sense of safety when deploying LLMs in high-stakes domains.
This paper studies tool-calling in large language model (LLM) agents, examining its effectiveness and efficiency. It analyzes evaluation pipelines, showing results are sensitive to implementation choices, and identifies computational waste in reinforcement learning training.
Multilingual large language models (MLLMs) often exhibit inconsistent behavior regarding cultural identity when the prompt's language changes. Researchers introduce a new metric, Singleton Fleiss's "k_S", and a consensus-driven alignment framework, C-3PO, to mitigate these cross-lingual cultural inconsistencies, achieving significant improvements.
ToolWeave is a structured framework for synthesizing realistic multi-turn tool-calling dialogues, essential for LLMs to function as autonomous agents. It addresses challenges in existing synthetic data generation by supporting realistic multi-step workflows and reducing parameter hallucination.
This content details the creation of a 24/7 autonomous AI agent system on a $6/month VPS, leveraging OpenClaw, DeepSeek V4 Pro, and Playwright for automation. The system manages social media posts, Dev.to articles, and a Gumroad store, showcasing cost-effective and efficient AI automation.