Tag: #agentic-reasoning5 results
SCALAR introduces an Actor-Critic-Judge pipeline for AI-assisted reasoning in quantum field theory and string theory. Multi-turn dialogues outperform single-shot attempts, with improvements varying by Actor-Critic pairing and feedback strategies. It provides a testbed to evaluate interaction structures for AI-driven scientific discovery.
Meta's semi-formal reasoning prompting requires LLMs to complete logical certificates by stating premises, tracing execution paths, and deriving conclusions for execution-free code analysis. It achieves up to 93% accuracy in code review, fault localization, and QA tasks. This cuts infrastructure costs for agentic code reasoning in large repos.
GEARS reframes large-scale ranking as an agentic discovery process in a programmable environment, addressing engineering bottlenecks over modeling limits. It encapsulates expert knowledge into reusable agent skills for high-level intent steering and includes validation hooks for robust, stable policies. Experiments across product surfaces show superior near-Pareto-efficient outcomes.
Apple introduces AMUSE, a benchmark for multimodal LLMs in agentic multi-speaker dialogue scenarios. Models like GPT-4o and Qwen3-Omni excel in perception but falter in tracking speakers, roles, and events across audio-video streams. It targets applications like conversational assistants and meeting analytics.
SideQuest uses the Large Reasoning Model (LRM) itself to compress KV cache by reasoning about token usefulness in long-context agentic tasks. It runs compression as a parallel auxiliary task to avoid polluting the main reasoning context. Evaluations with a model trained on 215 samples show 65% peak token reduction with minimal accuracy loss, beating heuristics.