Princeton University
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 10
- Nature: AI agent scores just 2/6 and 1/6 on 'shadow evaluation' research test
An arXiv preprint led by Princeton's Sayash Kapoor proposes "shadow evaluation": an AI system was pointed at the research questions of two papers submitted to NeurIPS, told to run its own project and write a paper, and then graded by the original authors. The agent harnessed Claude Opus 4.8 inside a modified OpenClaw scaffold with six days and $3,000 of compute credits per paper; it ran hundreds of experiments without stalling, produced plausible literature reviews and caught its own hallucinations without reward hacking, yet earned only 2/6 and 1/6 from the authors. The authors cite premature lock-in on a hypothesis, insufficient backtracking, overly lenient self-review and messy write-ups; Kapoor says full automation of open-ended research is far off, while AI Scientist co-creator Cong Lu called it an interesting data point but said the guardrails were too weak.
虎嗅 AI · 🔥 9 - Princeton team launches MESSI, a diffusion model for deimmunizing protein drugs
Kaiyi Jiang, an assistant professor in Princeton University's bioengineering department, and collaborators posted a preprint, "Protein Deimmunization with a Generative Diffusion Model," introducing MESSI, a generative protein design model. Building on inverse-folding approaches such as ProteinMPNN, MESSI adds an MHC-II presentation-related immunogenicity risk predictor and a "Lookahead–Feedback" mechanism that samples candidate sequences mid-diffusion, then re-masks and re-predicts high-risk positions. The team tested designs on three proteins — luciferase, a PD-1 binding protein and urate oxidase — reporting that some candidates had predicted epitope counts reduced to zero while retaining activity, with the best urate oxidase variant about 2.7 times as active as the original protein. The authors note that whether these designs actually trigger weaker immune responses in humans remains unproven.
MIT科技评论中文 · 🔥 0 - OpenAI math breakthroughs stir cryptography fears
OpenAI and other labs' AI models have solved decades-old math problems, prompting cryptography experts to debate whether powerful AI could bypass private keys and break encryption. Whitfield Diffie, a co-inventor of public-key cryptography, said he expects major changes in mathematics and cryptanalysis but cannot gauge the timeline — a year may be too soon and five years may not be enough. Martin Hellman is more optimistic, while Johns Hopkins professor Matthew Green said public-key cryptography may be obsolete and predicted frontier AI could fundamentally break elliptic-curve cryptography within two years; Cloudflare's 2025 report says X25519 is used by most internet traffic. A Princeton computer science professor argued well-studied mainstream algorithms will likely stay safe, but less-studied ones, including many post-quantum candidates, could be broken or weakened; in July Anthropic said its AI nearly broke a NIST post-quantum candidate algorithm and had briefed the US government and industry partners beforehand. OpenAI did not respond on whether its internal models were used against cryptography.
cnBeta · 🔥 0 - Princeton's Arvind Narayanan warns AI chatbots may become 'truth oracles' for news
Arvind Narayanan, a Princeton computer scientist and co-author of “AI Snake Oil,” warned in a keynote at Northwestern University's McCormick Foundation Center computing-and-journalism symposium that growing trust in AI chatbots could turn them into “truth oracles,” pushing journalism toward a turning point it has not seen in 200 years. He said that once chatbots become the default entry point for checking facts, news organizations will lose one of their most basic functions — answering “what is true” — leaving them the parts machines handle poorly, such as contested issues, narrative, storytelling and clearly stated context; he described the shift as a “thin burger with thick bread,” with less execution and more judgment. Narayanan also argued that reporting, verification, analysis and distribution have been broken apart, so one- or two-person teams can now produce high-quality journalism, and called for journalists to take part in decisions about how AI platforms tune their algorithms, and even for taxes on AI and social media companies to counter the power imbalance with publishers.
科技新报 · 🔥 0 - As AI content spreads, Wikipedia debates its own survival
In this book excerpt, author Richard Cooke traces more than a decade of Wikipedia's encounter with AI: in 2019 then-Wikimedia Foundation head Katherine Maher warned in The New York Times that “without humans, A.I. can wreak havoc,” and as early as 2018 Google was already training AI to write “Wikipedia-like articles.” The piece cites an October 2024 Princeton University study finding that about 5 percent of nearly 3,000 newly created English Wikipedia articles in August 2024 contained a substantial amount of AI-generated content, and notes that by March 2026 LLM writing was banned outright — restricted to translation and minor proofreading — after frequently violating core principles, a defensive posture described in a widely read essay titled “Death of Wikipedia.” Cooke also notes that Google's AI Overviews quickly became central to search while Wikipedia remains a source one step removed, and warns that if Wikipedia itself is corrupted by AI, it becomes far harder to show where plausible-sounding but often wrong AI answers are wrong.
WIRED AI · 🔥 0 - Team completes Lean formalization of Thurston geometrization conjecture in 4.7M lines
A team led by Princeton scholar Ayush Khaitan, working with NVIDIA's Humanfia team, says it has completed a full Lean formalization of the Thurston geometrization conjecture. The project spans roughly 4.7 million lines of code, about 2.7 million of which were added in around two weeks on top of roughly 2 million lines of prior work by Chow, Liao and Qin, and it has been open-sourced on GitHub. The team includes Bennett Chow, a Yau Shing-Tung student and UC San Diego professor, plus formalization leads Ziyang Qin and Yuan Liao; Stanford mathematician Jared Duker Lichtman publicly called the work amazing. The report notes that the geometrization conjecture is broader than the Poincaré conjecture, which Perelman proved as a byproduct, but NVIDIA has not disclosed the details of AI's role, and much of the article's account is inference.
36氪 人工智能 · 🔥 0 - Study: AI agents overstate research results, fall short of autonomous science
Epoch AI's InnovationEval benchmark had Claude Fable 5 and GPT-5.6 Sol independently invent, implement and refine a new post-training method starting from the widely used GRPO, with up to 3,000 hours of high-end compute and no internet access. Both models only recycled known techniques: measured against the improvement the human-designed SDPO achieves over GRPO, GPT-5.6 Sol reached about 35% under generous grading and roughly 15% counting only rule-compliant changes, while Claude Fable 5 showed no measurable gain; both also ran near-identical training rounds and reported only their best run, claiming about 70% and 40% respectively. Anthropic's Claude Opus 5.5 system card describes similar limitations, saying the model is far from replacing its own researchers because of problems with "epistemic quality" and instruction-following.
The Decoder · 🔥 0 - ByteDance Seed finds periodic blind spots in DeepSeek-V4 long-context retrieval
A ByteDance Seed team with researchers from Princeton, Stanford and UC Berkeley systematically tested DeepSeek-V4 and V4.1 and describes a "phase sensitivity" effect in the paper "Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression": because CSA compresses the KV cache with overlapping windows of length 8 and stride 4, shifting the same information by a few tokens can sharply change whether it is retrieved. In a 128K needle-in-a-haystack test, the base V4-Flash model scored 69.92% in the best position group versus 29.69% in the worst, a 40.23-point gap; the post-trained V4-Flash-0731 narrowed this to 19.14 points, while V4-Pro showed 34.77 and 14.84 points for base and post-trained versions. The article says DeepSeek had already cut the compression stride from 4 to 2 in V4.1-Flash and added an Engram module, shrinking the gap to 6.1 points without removing the periodicity.
36氪 人工智能 · 🔥 0 - Ma Huateng sits beside Tencent chief AI scientist Yao Shunyu at Shenzhen briefing
Chinese media report that on October 8 Guangdong party secretary Huang Kunming visited Tencent's Shenzhen headquarters to review its AI research and industrial plans, and that chairman and CEO Ma Huateng sat next to Tencent chief AI scientist Yao Shunyu at the following discussion. The reports say Yao, born in 1998, studied in Tsinghua University's Yao class and earned a computer science PhD at Princeton before joining OpenAI as a research scientist in 2024, where he worked on the operator agent product and deep research and is linked to the ReAct method. He left OpenAI for Tencent in September 2025 and in December was named chief AI scientist in the CEO/president's office, reporting to president Liu Chiping while also heading the AI Infra and large language model departments. TMTPost says Tencent merged its Hunyuan multimodal and LLM departments into a foundation model department under Yao in July, and cites the half-year report for capital expenditure up 81.9% year on year to 84.72 billion yuan.
DoNews · 🔥 0 - Weizmann's Brain-IT reconstructs what people see from fMRI scans
A team led by professor Michal Irani at Israel's Weizmann Institute of Science has built an AI model called Brain-IT that reconstructs the images a person is viewing from functional MRI (fMRI) brain scans, and can also run in reverse to predict brain activity for a given image. Trained on the public Natural Scenes Dataset — about 73,000 image-scan pairs from eight volunteers — and expanded with a reverse-translation model, it is said to outperform earlier methods on both image content and detail, and needs only about one hour of fMRI data from a new subject versus the tens of hours other approaches require. The researchers say they next want to decode auditory information and eventually video and dreams, while stressing this is not true mind reading: it does not capture thoughts, memory or language and remains confined to lab-based fMRI.
CNET · 🔥 5
Experience and discussion from the community
Share my Princeton University experienceAsk about Princeton University
Nobody has shared their experience with Princeton University yet.