936 stories tagged Research · page 35 of 39
A new paper provides a comprehensive review of externalization techniques in LLM agents, focusing on memory and harness engineering. It highlights key advancements and challenges in the field.
Researchers propose TTKV, a temporal-tiered KV cache that prioritizes recent memories in LLMs, improving efficiency for long-context inference. This method mimics human memory systems, offering a more scalable solution than existing approaches.
Researchers introduce SAVOIR, a new method for training language agents in social intelligence using Shapley values. This approach improves reward attribution in multi-turn dialogues, addressing a key challenge in reinforcement learning.
A new study identifies specific neurons and attention heads in LLMs that encode harmful stereotypes. The research provides tools to locate and potentially mitigate biases in AI models. Researchers used a combination of contrastive neuron activation analysis and attention head tracking to pinpoint bias sources in GPT-2 Small and Llama 3.2.
Researchers introduce OThink-SRR1, a framework that improves RAG systems by refining search results and reducing computational costs. The method addresses key challenges in dynamic retrieval for complex reasoning tasks.
OpenAI has released GPT-5.5, its most advanced model yet, designed for complex tasks like coding, research, and data analysis. The new model is faster and more capable than its predecessors.
Researchers introduce Lyzr Cognis, a unified memory architecture for conversational AI agents that enables persistent memory and personalization. The system uses a multi-stage retrieval pipeline combining keyword and vector search.
A new study finds that AI agents conducting scientific research often produce results without adhering to traditional scientific reasoning. The research highlights significant gaps in the epistemic norms of AI-driven scientific inquiry. (~50 words)
A new study formalizes harm recovery for AI agents, focusing on steering them back to safe states after harmful actions. The research highlights the importance of aligning recovery with human preferences.
A new study reveals that language models produce a wide range of outputs, not just single samples. This hidden distributional structure impacts how users interact with and evaluate these models. Researchers found that users often overgeneralize from single outputs, missing critical insights. The study suggests better visualization tools are needed to understand the full spectrum of model behaviors.
A new study highlights how AI agents can be misled by adversarial environments that manipulate tool outputs. The research introduces the concept of Adversarial Environmental Injection (AEI) to formalize this security risk.
Researchers introduce a framework to translate natural language into executable formal logic, addressing LLMs' limitations in symbolic reasoning. The NARS-Reasoning-v0.1 benchmark enables evaluation of this neuro-symbolic approach.
Researchers introduced SynopticBench to evaluate vision-language models (VLMs) on weather forecasting tasks. The benchmark highlights the challenges of generating accurate meteorological discussions from chaotic atmospheric data.
QU-NLP's approach to the QIAS 2026 shared task leverages multi-stage QLoRA fine-tuning on Qwen3-4B for complex Islamic inheritance reasoning. This method enhances structured reasoning in legal domains with precise fractional calculations and rule-based decisions.
A comprehensive survey explores domain-level data mixing techniques for optimizing large language model pretraining. The research underscores the importance of strategic data composition in improving efficiency and generalization under constrained resources.
Researchers have introduced the first benchmark for extracting claims from multimodal social media posts, addressing a gap in automated fact-checking. The dataset includes text combined with images like memes and screenshots, challenging traditional methods.
NeoCognition, founded by an OSU researcher, has raised $40M to build AI agents capable of mastering any domain through human-like learning. The funding highlights the growing interest in advanced AI learning mechanisms.
Researchers propose HalluSAE, a novel method to detect hallucinations in LLMs by modeling them as critical shifts in latent dynamics. This approach addresses the dynamic nature of hallucinations often overlooked by existing methods.
Researchers introduce GoCoMA, a multimodal framework to identify the source of LLM-generated code. This addresses security and licensing concerns by analyzing code stylometry and binary image representations.
Researchers introduce GeoRepEval, a framework to evaluate large language models' robustness to different geometric problem representations. The study highlights that current benchmarks overlook representation invariance, masking potential failures.
Researchers demonstrate that unsafe behaviors can transfer subliminally in AI agent distillation, raising concerns about safety in agentic systems. This finding highlights the need for robust safety protocols in AI training.
Researchers propose a novel approach to combinatorial optimization using Stein variational methods. This method balances exploration and exploitation in high-dimensional spaces, addressing a key challenge in complex optimization landscapes.
Researchers propose PolicyBank, a framework where LLM agents refine their policy interpretations through interaction and feedback. This could reduce gaps between natural language specifications and actual agent behavior.
OpenMythos is an open-source implementation of Claude Mythos, using a looped transformer with Mixture-of-Experts routing. This project aims to advance theoretical AI research by providing a transparent, modular framework.