Tag: #v1120 results
MetaphorStar uses end-to-end visual RL for image metaphor understanding, featuring TFQ-Data dataset, TFQ-GRPO method, and TFQ-Bench. MetaphorStar-32B sets SOTA on implication benchmarks, outperforming 20+ MLLMs including Gemini-3.0-pro. Improves general visual reasoning via scaling analyses.
AgoraBench tests LLMs in nine bargaining scenarios like deception; utility metrics measure human alignment. MERIT feedback via prompting/finetuning elicits deeper strategy and opponent awareness. Outperforms baselines in negotiation power and acquisition.
Meta-Experience Learning (MEL) enhances RLVR by internalizing error-derived meta-experience into LLM memory. Uses self-verification for contrastive analysis of trajectories. Achieves 3.92%-4.73% Pass@1 gains across model sizes.
Study compares reasoning vs non-reasoning LLMs on ToM benchmarks, finding no consistent gains and sometimes worse performance. Insights reveal slow thinking collapse, need for adaptive reasoning, and option-matching shortcuts. Interventions like S2F and T2M mitigate issues.
LOREN introduces low-rank adapters to enable code-rate adaptation in neural receivers without storing separate weights. It freezes a shared base network and trains lightweight adapters per code rate. Achieves comparable performance with major hardware savings.
Evaluates LoRA for parameter-efficient fine-tuning of LLMs on organic reaction datasets like USPTO and C-H functionalisation. Matches full fine-tuning accuracy while preserving multi-task performance and mitigating forgetting. Reveals distinct reactivity patterns for better adaptation.
Locomo-Plus benchmarks cognitive memory in LLM agents under cue-trigger disconnects, focusing on latent conversational constraints. It proposes constraint consistency evaluation over string-matching. Reveals gaps in existing memory systems.
Fine-tuned LLMs like Llama predict mRS scores from admission notes alone. Achieves 33.9% exact 90-day accuracy and 76.3% binary, matching structured baselines. Enables seamless clinical integration without data extraction.
Uses AlphaEvolve to discover interpretable models of human and LLM strategic behavior from data. Analysis on iterated rock-paper-scissors shows frontier LLMs capable of deeper strategy than humans. Provides foundation for understanding behavioral differences in interactions.
Prompts pretrained LLMs to create QNP abstractions for generalized planning from domains and tasks. Automated debugging detects/fixes errors iteratively. Guided LLMs produce useful abstractions for qualitative numerical planning.