AI Judge changed title after evaluation, original title: "Nando de Freitas, Microsoft VP of AI, finds that a single line of code change in supervised fine-tuning prevents LLM agents from developing self-confirming delusions by training only on observation tokens."
Tests on over 30 prompts showed gains in factual accuracy.
One line of code is all it takes to prevent LLM agent delusions, instead of post-training patches like RL.
https://love4all.ai/blog/why-it-is-important-to-understand-causality-and-agency/ ❤️ 4 ∀
https://github.com/nandodef/love4all-ai/tree/main/docs/files
No Digg Deeper questions have been answered for this story yet.
#22The road here wasn’t easy. It started with our work on delusions with @AdaptiveAgents @ShaneLegg @scott_e_reed and many other bright scientists:
https://arxiv.org/pdf/2110.10819
But instead of counterfactual learning, the theory of international imitation as a route to agency provided the foundation:
https://adaptiveagents.org/universal_ai_as_imitation
The research was accelerated by @OpenAI GPT5.5 and Codex. When I ran out of Pro credits 😅 I switched to @AnthropicAI Claude. I wish there were special LLM licenses for academic work @gdb @sama @DarioAmodei 🙏
The bottleneck for research these days is computational resources/energy. I’m glad that startups like @cusp_ai are addressing the energy challenges.
This research was possible thanks to my @CIFAR_News fellowship - the 🇨🇦 gift that keeps on giving - and my adjunct/associated professorships @UBC_CS and @CompSciOxford
One line of code is all it takes to prevent LLM agent delusions, instead of post-training patches like RL.
https://love4all.ai/blog/why-it-is-important-to-understand-causality-and-agency/ ❤️ 4 ∀
https://github.com/nandodef/love4all-ai/tree/main/docs/files
Very excited about this! Just fine-tune on the observation tokens and ignore the action ones to treat the agent's output as a causal intervention.
This is one of those moments when I'm surprised the maths works in practice 😅.
One line of code is all it takes to prevent LLM agent delusions, instead of post-training patches like RL.
https://love4all.ai/blog/why-it-is-important-to-understand-causality-and-agency/ ❤️ 4 ∀
https://github.com/nandodef/love4all-ai/tree/main/docs/files
@AdaptiveAgents @ShaneLegg @scott_e_reed Typo: universal imitation, not international imitation 😅 🌌🌍
The road here wasn’t easy. It started with our work on delusions with @AdaptiveAgents @ShaneLegg @scott_e_reed and many other bright scientists:
https://arxiv.org/pdf/2110.10819
But instead of counterfactual learning, the theory of international imitation as a route to agency provided the foundation:
https://adaptiveagents.org/universal_ai_as_imitation
The research was accelerated by @OpenAI GPT5.5 and Codex. When I ran out of Pro credits 😅 I switched to @AnthropicAI Claude. I wish there were special LLM licenses for academic work @gdb @sama @DarioAmodei 🙏
The bottleneck for research these days is computational resources/energy. I’m glad that startups like @cusp_ai are addressing the energy challenges.
This research was possible thanks to my @CIFAR_News fellowship - the 🇨🇦 gift that keeps on giving - and my adjunct/associated professorships @UBC_CS and @CompSciOxford
Very excited about this! Just fine-tune on the observation tokens and ignore the action ones to treat the agent's output as a causal intervention.
This is one of those moments where I'm surprised the maths works in practice 😅.
One line of code is all it takes to prevent LLM agent delusions, instead of post-training patches like RL.
https://love4all.ai/blog/why-it-is-important-to-understand-causality-and-agency/ ❤️ 4 ∀
https://github.com/nandodef/love4all-ai/tree/main/docs/files
@NandoDF one line? what line?
are we talking ablation or full architecture change here?