Dec 31, 2025 · 34m · latent-space
[State of AI Papers 2025] Fixing Research with Social Signals, OCR & Implementation — Team AlphaXiv
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the Latent Space podcast, the founders of AlphaXiv discuss transforming static academic papers into interactive, executable research artifacts while analyzing standout NeurIPS architectures and practical AI solutions for the peer review and reproducibility crisis.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Rayhan rejects the host's critique by arguing AlphaXiv is not trying to solve universal academic reproducibility, but rather catering strictly to applied engineers via power-law implementation ease.
Hardest push from the hosts ▶ 31:19 Host warns founders they are biting off more than they can chewHost explicitly pushes back against the company's product direction, warning that building executable sandboxes for papers is an impossible task that caused Replicate to pivot.
Biggest teaching moment ▶ 11:36 Rayhan breaks down Tiny Recursive ModelsRayhan explains the mechanics of 7-million parameter recursive latent transformers achieving high ARC-AGI scores with minimal compute compared to massive frontier reasoning models.
The host holds their own ▶ 32:24 Host cites Brev and NVIDIA Launchables infrastructureHost demonstrates deep domain expertise by citing NVIDIA Launchables and his background as an investor in Brev, explaining domain-specific solutions for hosting GPU environments.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Origins of AlphaXiv at Stanford | 2 | 2 | 1 | 1 | Host opens by inquiring about AlphaXiv's origins. The founders share their Stanford dorm background and the student research frustrations that led to building paper annotations. | |
| Differentiating from Hugging Face and Exploring OCR Models | 6 | 4 | 2 | 3 | Host probes why AlphaXiv won over Hugging Face and drills into PDF parsing and OCR model options. Rayhan details DeepSeek OCR's price-to-performance on A100s versus API alternatives. | |
| Expanding Beyond Papers to Interactive Research Artifacts | 5 | 4 | 2 | 3 | Host asks if AlphaXiv aims to replace Papers with Code or act as a recommendation system. The team outlines expanding from PDFs to executable Docker implementations. | |
| Favorite Papers at NeurIPS: Compute-Efficient Architectures | 5 | 6 | 1 | 2 | Rayhan shares compute-efficient NeurIPS papers including Tiny Recursive Models and low-rank evolutionary strategies. Host asks clarifying questions on search versus gradient descent. | |
| AI for Science: Agent Laboratory and Virtual Labs | 6 | 4 | 3 | 6 | Host pushes back against grouping all automated research under 'science' and questions recurring automated scientist claims. Guests reframe AI agents as pragmatic virtual lab assistants. | |
| Reinforcement Learning for Reasoning and the Rise of Qwen | 6 | 5 | 2 | 4 | Co-founders discuss Agent R-One and sample-efficient Qwen fine-tuning. Host references the presence of the Qwen team at NeurIPS and compares Qwen's trajectory with DeepSeek. | |
| The Academic Peer Review Crisis and AI Linters | 6 | 4 | 2 | 3 | Host raises the breakdown of academic peer review at conferences like ICLR. Founders discuss their refusal to build paper-writing bots while exploring pre-submission AI linters. | |
| Overcoming Semantic Search with Social Signals and Dynamic Artifacts | 5 | 5 | 2 | 2 | Rayhan explains why naive semantic search fails across 3 million arXiv papers without social signals. Host brings up Emergent Mind using YouTube traction as an alternate signal. | |
| Empowering Applied Engineers and Ranking Implementation Ease | 7 | 5 | 3 | 7 | Host gives direct pushback warning that automating paper execution environments is an impossible task, citing Replicate's pivot. Rayhan counters that they only focus on power-law papers and author-assisted setup. | |
| Conclusion and Future Outlook | 1 | 0 | 0 | 0 | Host wraps up the interview, commends AlphaXiv's mission, and gives closing well-wishes. |
Statements from this episode (0)
Nothing in this episode matches those filters. clear them