# NeurIPS 2025 Best Paper Awards

## NeurIPS 2025 awarded seven papers
- **NeurIPS 2025** awarded seven papers for their significant contributions to machine learning, covering topics like **diffusion models**, **self-supervised reinforcement learning**, and **attention mechanisms** in large language models, showcasing the conference's commitment to advancing AI research.

## State of AI: An Empirical 100T Token Study with OpenRouter
- The **State of AI** report from OpenRouter highlights significant advancements in **machine learning** and **natural language processing**, showcasing how these technologies are reshaping industries and enhancing productivity. [Read more here](https://openrouter.ai/state-of-ai).

## Gemini 3 Pro: the frontier of vision AI
- **Gemini 3 Pro** is Google's most advanced multimodal model, excelling in **document, spatial, screen, and video understanding**, enabling complex visual reasoning and document processing.

## CUDA-l2: Surpassing cuBLAS performance for matrix multiplication through RL
- **CUDA-L2** leverages **reinforcement learning** to optimize Half-precision General Matrix Multiply (HGEMM) CUDA kernels, achieving superior performance compared to established libraries like **cuBLAS** and **torch.matmul** across 1,000 configurations on the A100 GPU.

## Addressing the “LLMs with thousands of tools” challenge
- **Anthropic's _Tool Search_ feature** aims to address the **“too many tools in context”** issue by enabling models to discover tools on-demand, rather than preloading thousands of definitions.

## Reframing Impact
- **An impact measure could serve as the first safeguard against powerful agents with imperfect objectives, enabling effective control without needing to define the objective itself.** This approach emphasizes the importance of creating a robust framework that can mitigate risks associated with advanced AI systems.

## Visualizing emergent structure in the Dragon Hatchling (BDH)
- The **BDH architecture** offers a novel approach to pathfinding by modeling **neuron-to-neuron interactions** on sparse graphs, utilizing **Hebbian learning** to adapt its circuits dynamically, which distinguishes it from traditional transformer models.

## Mitigating Object and Action Hallucinations in Multimodal LLMs via Self-Augmented Contrastive Alignment
- The **SANTA framework** effectively mitigates **object and action hallucinations** in multimodal LLMs by focusing on **visual facts** and reducing spurious correlations, enhancing the accuracy of video descriptions.

## Tiny Recursive Models (TRMs), Hierarchical Reasoning Models (HRMs)
- **Tiny Recursive Models (TRMs)** leverage recursion to perform extensive computations with **fewer parameters**, enhancing efficiency in hierarchical reasoning tasks.

## Semantic Soft Bootstrapping: Long Context Reasoning in LLMs without Reinforcement Learning
- **Semantic Soft Bootstrapping (SSB)** enhances long context reasoning in large language models (LLMs) by utilizing a self-distillation technique that allows the model to act as both teacher and student, improving its reasoning capabilities without the need for reinforcement learning.

## Multi-LLM Collaboration for Medication Recommendation
- **Multi-LLM collaboration** enhances medication recommendations by addressing the **hallucination** and **inconsistency** issues prevalent in individual large language models (LLMs), utilizing a Chemistry-inspired framework for improved reliability.

## Arbitrage: Efficient Reasoning via Advantage-Aware Speculation
- **Arbitrage** introduces a dynamic routing mechanism that enhances **step-level speculative generation**, significantly improving the efficiency of reasoning tasks in Large Language Models by reducing unnecessary rejections of semantically equivalent steps.

## SEASON: Mitigating Temporal Hallucination in Video Large Language Models via Self-Diagnostic Contrastive Decoding
- **SEASON** introduces a novel method, **Self-Diagnostic Contrastive Decoding**, to combat **temporal hallucination** in Video Large Language Models (VideoLLMs), enhancing both temporal and spatial accuracy without requiring additional training.

## AdmTree: Compressing Lengthy Context with Adaptive Semantic Trees
- **AdmTree** introduces a **novel framework** for adaptive context compression in Large Language Models (LLMs), focusing on **high semantic fidelity** while enhancing computational efficiency through a hierarchical structure.

## TaskEval: Synthesised Evaluation for Foundation-Model Tasks
- **TaskEval** introduces a novel **synthesised evaluator** for Foundation Model (FM) tasks, addressing the critical issue of **hallucinations** in applications by automating evaluation and integrating human feedback.