Large Language Models

ResearchBench: Benchmarking LLMs for Scientific Discovery via Inspiration-Based Task Decomposition.

This survey presents a methodology-centered taxonomy of LLM agent systems, linking design principles to emergent behaviors and identifying future research directions.

OlymMATH: A new Olympiad-level math benchmark rigorously tests LLMs’ reasoning, revealing limitations and paving the way for advancements.

Survey on improving efficiency in large reasoning models across language, multimodality, and beyond.

Boost LLM reasoning by having models ‘Think Twice’! This novel method iteratively refines answers, significantly enhancing accuracy on complex tasks.

FFN Fusion: Parallelizing sequential computation in large language models for significant speedups!

LLMs falter on culturally adapted math problems, revealing a critical cultural bias.

V-SEEK accelerates LLM reasoning on open-hardware RISC-V platforms, achieving up to 3.0x speedup through optimized kernels and memory management.

MARS: Optimizing prompts with multi-agent collaboration and Socratic learning for better LLM performance!

LEMMA: LLMs learn math via mistake analysis and correction, boosting performance without external critics.

XAttention: Antidiagonal scoring unlocks block-sparse attention, slashing compute costs in long-context Transformers without sacrificing accuracy.

A comprehensive survey on evaluation methodologies for LLM-based agents, analyzing benchmarks and frameworks across key dimensions like capabilities, applications, and generalist performance.

LLMs survey: Model, output, and prompt-based strategies for efficient reasoning, mitigating ‘overthinking’ for faster, cheaper, and real-world applications.

MathFusion: Instruction Fusion enhances LLM’s math problem-solving!

CaKE: Editing LLMs to Enhance Knowledge Generalization Across Reasoning Tasks.

BIGO(Bench) can help LLMs generate code with controlled time/space complexity, addressing the gap in current evaluations and encouraging further exploration.

A new test-time method, Temporal Consistency, is introduced to improve LLM reasoning by leveraging iterative self-reflection.

Pensez: Strategic fine-tuning beats massive data for superior reasoning in French LLMs, challenging conventional wisdom.

Φ-Decoding: Adaptive foresight sampling balances inference-time exploration and exploitation for better LLM reasoning.

This research explores how well LLM uncertainty measures align with human uncertainty, finding Bayesian and top-k entropy measures show promise.