Paper Feed: April 2025

Evals

• BaxBench: Can LLMs Generate Correct and Secure Backends?
  Mark Vero, Niels Mündler, Victor Chibotaru, [...], Martin Vechev (2025)
• CVE-Bench: A Benchmark for AI Agents' Ability to Exploit Real-World Web Application Vulnerabilities
  Yuxuan Zhu, Antony Kellermann, Dylan Bowman, [...], Twm Stone, Daniel Kang (2025)
• DeltaBench (Can LLMs Detect Errors in Long CoT Reasoning?)
  Yancheng He, Shilong Li, Jiaheng Liu, [...], Wenbo Su, Bo Zheng (2025)
• Writing as a testbed for open ended agents
  Sian Gooding, Lucia Lopez-Rivilla, Edward Grefenstette (2025)
• EvalTree: Profiling Language Model Weaknesses via Hierarchical Capability Trees
  Zhiyuan Zeng, Yizhong Wang, Hannaneh Hajishirzi, Pang Wei Koh (2025)

Science of DL / Interp

• How do language models learn facts? Dynamics, curricula and hallucinations
  Nicolas Zucchet, Jörg Bornschein, Stephanie Chan, Andrew Lampinen, Razvan Pascanu, Soham De (2025)
• ARBOR
  Bau Lab (2025)
• Prospects for Alignment Automation: Interpretability Case Study
  Jacob Pfau, Geoffrey Irving (2025)
  
  Why this is notable
  Makes a nice argument: interpretability is automation-tractable. Namely, given a doubling time for AI R&D (informed by the METR work), and assuming we will be able to automatically verify interpretability progess (most notably via downstream tasks where interpretability methods improve in time complexity over behavioral methods), interpretability will be automated in the next ~5 years. Some scattered points: tasks must be robustly verifiable; the reward is defined via FLOP reduction+performance improvement; eg for component modeling g is I(M).
• Error Discovery by Clustering Influence Embeddings
  Fulton Wang, Julius Adebayo, Sarah Tan, Diego Garcia-Olano, Narine Kokhlikyan (2023)

Compute / Reasoning

• Most AI value will come from broad automation, not from R&D
  Ege Erdil, Matthew Barnett
• Gemstones: A Model Suite for Multi-Faceted Scaling Laws
  Sean McLeish, John Kirchenbauer, David Yu Miller, [...], Tom Goldstein (2025)
• Pro-1
  Michael Hla (2025)

General Safety

• Learning Task Decomposition to Assist Humans in Competitive Programming
  Jiaxin Wen, Ruiqi Zhong, Pei Ke, Zhihong Shao, Hongning Wang, Minlie Huang (2024)
• Unnatural Languages Are Not Bugs but Features for LLMs
  Keyu Duan, Yiran Zhao, Zhili Feng, [...], J. Zico Kolter, Michael Qizhe Shieh
• A Complete List of All (arXiv) Adversarial Example Papers
  Nicholas Carlini (2019-current)
  
  Why this is notable
  Recently discovered that Carlini's list isn't as widely known as I thought it was. This is a shame, because it is a great resource and probably worth skimming every couple of weeks if you are in the field.

Security/Control

• Trusted Machine Learning Models Unlock Private Inference for Problems Currently Infeasible with Cryptography
  Ilia Shumailov, Daniel Ramage, [...], Eugene Bagdasarian (2025)
• A Framework for Evaluating Emerging Cyberattack Capabilities of AI
  Mikel Rodriguez, Raluca Ada Popa, Four Flynn, Lihao Liang, Allan Dafoe, Anna Wang (2025)
• Unpatched AI: Security and Vulnerability Reports
  Unpatched AI (2025)

Miscellaneous

• Undermining Mental Proof: How AI Can Make Cooperation Harder by Making Thinking Easier
  Zachary Wojtowicz, Simon DeDeo (2025)
  
  Why this is notable
  I found this framework very useful. For example, I think this kind of set-up can mostly explain why (current) workshop paper submissions from AI-scientists are damaging to the field. The story is something like: the reviewing bar for workshops is quite low and submitting a paper is mainly a signal / proof-of-work. At a minimum, a workshop paper is a proof-of-mental-work that the author is thinking somewhat coherently about a topic. AI scientists undermine this signal by circumventing the (not very robust) workshop review process.
• attention is logarithmic, actually
  spike (2025)