Runjin Chen

Member of Technical Staff @AnthropicAI | PH.D. student @UTAustin @VITAGroupUT | Previously BS/MS @sjtu1896

[Runjin Chen @RunjinChen]

New Anthropic Research: Persona Vectors

We can:
1. Monitor how a model’s personality is changing during a conversation, or over training
2. Mitigate undesirable persona shifts during development or prevent during training.
3. Identify training data that leads to shift

[Anthropic @AnthropicAI]

New Anthropic research: Persona vectors.

Language models sometimes go haywire and slip into weird and unsettling personas. Why? In a new paper, we find “persona vectors"—neural activity patterns controlling traits like evil, sycophancy, or hallucination.

[Jack Lindsey @Jack_W_Lindsey]

Before limited-releasing Claude Mythos Preview, we investigated its internal mechanisms with interpretability techniques. We found it exhibited notably sophisticated (and often unspoken) strategic thinking and situational awareness, at times in service of unwanted actions. (1/14)

[Anthropic @AnthropicAI]

Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software.

It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans.
https://t.co/NQ7IfEtYk7

[Victor.Kai Wang @VictorKaiWang1]

very happy to release this parameter generation work.
from P-diff (2024), RPG (2025), DnD (2025) to HY-WU, parameter generation becomes more and more practical. thx @TencentHunyuan and happy to work with @oahzxl @mmbronstein @ZiqiaoWang63428 @cindy_x_wu @YangYou1991 @VITAGroupUT

[Junyuan "Jason" Hong @hjy836]

🧠 Conclusion: Data curation is cognitive hygiene for AI.
🩺 Regular data “health checks” are essential for keeping models reliable, safe, and aligned.

���� The striking parallels between AI and human cognitive decline may even offer new insights into human brain health.

👩‍🔬 Work by: Shuo Xing*, Junyuan Hong*, Yifan Wang, Runjin Chen, Zhenyu Zhang, Ananth Grama, Zhengzhong Tu, and Zhangyang Wang

🌐 Website (code, data, models): https://t.co/jLvQQRtSco
📄 ArXiv: https://t.co/i7bpjhlAUw

[Ethan Perez @EthanJPerez]

We’re hiring someone to run the Anthropic Fellows Program!

Our research collaborations have led to some of our best safety research and hires. We’re looking for an exceptional ops generalist, TPM, or research/eng manager to help us significantly scale and improve our collabs 🧵

[Runjin Chen @RunjinChen]

@infoxiao @giffmana However, we didn’t compare prompt-based methods with preventative steering during training. It might be worth exploring, for example, by always prepending an “evil” system prompt during training.

[Runjin Chen @RunjinChen]

@infoxiao @giffmana If you're referring to test time intervention, we actually compared two approaches in the appendix: using prompts to suppress undesirable personas versus using inference-time steering. We found that inference-time steering tends to be more effective.

[Runjin Chen @RunjinChen]

@infoxiao @giffmana I think persona vector goes beyond simple prompting, for instance, can be used to monitor personality changes during training or development.

[Emmanuel Ameisen @mlpowered]

In which the gang (@RunjinChen, @andyarditi, @Jack_W_Lindsey ):

- identifies vectors for bad personas (evil, sycophancy, hallucinations, etc)
- shows that if you inject the bad vectors in training, the model learns to not do the bad thing!!

aka vaccines but for LLMs

[Anthropic @AnthropicAI]

New Anthropic research: Persona vectors.

Language models sometimes go haywire and slip into weird and unsettling personas. Why? In a new paper, we find “persona vectors"—neural activity patterns controlling traits like evil, sycophancy, or hallucination.

[Victor.Kai Wang @VictorKaiWang1]

Customizing Your LLMs in seconds using prompts🥳!
Excited to share our latest work with @HPCAILab, @VITAGroupUT, @k_schuerholt, @YangYou1991, @mmbronstein, @damianborth : Drag-and-Drop LLMs(DnD). 2 features: tuning-free, comparable or even better than full-shot tuning.(🧵1/8)

[Runjin Chen @RunjinChen]

Our LLaGA excels in versatility, generalizability and interpretability, allowing it to perform consistently well across different datasets and tasks, extend its ability to unseen datasets or tasks, and provide explanations for graphs

[Runjin Chen @RunjinChen]

Thrilled to share our latest project, "LLaGA: Large Language and Graph Assistant"
🚀 Dive into our findings here: https://t.co/QhlOLHl8Mi.
Plus, access our code on GitHub: https://t.co/lbRU1BBqZW

[Runjin Chen @RunjinChen]

Key Feature: A versatile linear projector seamlessly bridges graph structures with the token space understood by Large Language Models (LLMs).