As we optimize model reasoning over verifiable objectives, how does this affect human understanding of said reasoning to achieve superior collaborative outcomes?. In our new preprint, we investigate human-centric model reasoning for knowledge transfer 🧵:

Thanks for reposting! I can’t believe I just noticed this 😓.

IMPersona is accepted at #COLM2025 + recommended for oral! Check out our work on imbuing human personality and memories into LLMs, allowing them to evade detection even by close friends and family. @COLM_conf.

RT @ori_press: Do language models have algorithmic creativity?. To find out, we built AlgoTune, a benchmark challenging agents to optimize….

This and lots more insights (like a trajectory visualizer) at Thanks to Carlos, Diyi, Nick, Shunyu, and Karthik for helping make this happen! I’ve wanted to do this project for an entire year and it’s so rewarding finally seeing it come to fruition :).

As we build more powerful AI, we need to be equally intentional about building AI that can effectively teach and collaborate with humans: otherwise we risk creating a world of powerful but incomprehensible AI assistants.

By clustering embeddings of user queries, feedback, and model responses, we see a nuanced picture of the interaction types that drive effective/non-effective knowledge transfer in collaboration.

Even with the overall positive correlation, the significant outliers (ie Claude-3.7-Sonnet) prove that some models are just better teachers relative to their capabilities, suggesting diminishing returns can be broken with intentional post-training.

This diminishing returns pattern suggests we could be heading toward a future where:. 1. AI capabilities race ahead of human understanding.2. The "explanatory debt" compounds over time.3. Humans become increasingly dependent on AI they understand less and less.

We obtain scaling trends of collaboration v reasoning. We find:. 1. The gap between "what AI can do" and "what AI can teach humans to do" may be widening as models improve.2. Humans don’t necessarily prefer stronger reasoners: preferences are domain and skill specific

We recruit 118 participants in a large-scale study + design a framework to isolate and measure knowledge transfer. Human ideate with models in an information exchange phase, then independently execute solutions, relying on internalization + combination of model + human insights.

How do we define knowledge transfer? We parameterize it as a model’s ability to project knowledge in model representational space, to human representational space: aka, its ability to convey to you what they know. (Adapted from @_beenkim)

RT @a1zhang: Can GPT, Claude, and Gemini play video games like Zelda, Civ, and Doom II?. 𝗩𝗶𝗱𝗲𝗼𝗚𝗮𝗺𝗲𝗕𝗲𝗻𝗰𝗵 evaluates VLMs on Game Boy & MS-DOS….

This and lots more insights here!.

During communication, we each possess a fundamental intent. However, the textual representation of this intent varies remarkably across individuals. Current prompting methods struggle to effectively override the deeply ingrained stylistic patterns, connotational nuances, and.

Now take a look at LLaMA-8B, which was fine-tuned on just 500 messages + also given my memories. Although the information it conveys isn’t very out of this world, it nails down the typical sort of apathetic and dryly sarcastic texting content that I normally dish out.

Let’s dive into some examples. This is GPT-4o pretending to be me, receiving my messaging data and memories through prompting. Sure, it accesses personal details correctly… but immediately draws suspicion due to how excited it was just to be there…

I gave both LLaMA-8B and GPT-4o access to my messages, and tasked them with pretending to be me, seeing if close friends and family could tell the difference. LLaMa was able to deceive them 44% of the time. GPT-4o only 6%. Why is this?!

RT @a1zhang: Claude can play Pokemon, but can it play DOOM?. With a simple agent, we let VLMs play it, and found Sonnet 3.7 to get the furt….

[9/9] 🙏 to collaborators @_carlosejimenez @karthik_r_n and our amazing testers! What began as a side project while waiting for lengthy human experiments evolved quickly. For case studies and detection strategies, check our full paper: Stay vigilant!.