Kaiqu Liang

🤔 Feel like your AI is bullshitting you? It’s not just you. 🚨 We quantified machine bullshit 💩 Turns out, aligning LLMs to be "helpful" via human feedback actually teaches them to bullshit—and Chain-of-Thought reasoning just makes it worse! 🔥 Time to rethink AI alignment.

🗓️ Mark your calendar: The 1st #ICRA Workshop on Public Trust in Autonomous Systems (PTAS) is just two days away! We'll explore the critical question: How do we build assurances into autonomous technologies from the ground up, shaping public trust before widespread deployment?

Awesome, I've been saying this for a while, inspired by @DrJohnVervaeke. LLMs are formally bullshitting, yes. medium.com/@balazskegl/on… A couple of threads that may be interesting: x.com/balazskegl/sta… x.com/NandoDF/status… The connection: when we speak, we have an…

Chain of thought can hurt LLM performance 🤖 Verbal (over)thinking can hurt human performance 😵‍💫 Are when/why they happen similar? Come find out at our poster at West-320 ⏰11am tomorrow! #ICML2025

LLMs are picking up weird patterns from humans. Aligning them to be helpful actually teaches them to bullshit? Must be why I like Grok Unhinged the best. :-)

arxiv.org/pdf/2507.07484 machine-bullshit.github.io Princeton University and UC Berkeley published a formalized analysis on the emergent dishonesty that rlhf at scale optimizes for in large language models, They provide a taxonomy and scoring system to allow for direct indexing,…

🚀 Excited to share the most inspiring work I’ve been part of this year: "Learning to Reason without External Rewards" TL;DR: We show that LLMs can learn complex reasoning without access to ground-truth answers, simply by optimizing their own internal sense of confidence. 1/n

40% with just 1 try per task: SWE-agent-LM-32B is the new #1 open source model on SWE-bench Verified. We built it by synthesizing a ton of agentic training data from 100+ Python repos. Today we’re open-sourcing the toolkit that made it happen: SWE-smith.

Introducing COMPACT: COMPositional Atomic-to-complex Visual Capability Tuning, a data-efficient approach to improve multimodal models on complex visual tasks without scaling data volume. 📦 arxiv.org/abs/2504.21850 1/10

@OpenAI rolled back GPT-4o citing sycophancy—just as our research predicted. Short-term feedback teaches AI to sound nice... and systematically misaligns it! Our solution: RLHS, training AI with simulated hindsight feedback for long-term alignment! 👉 arxiv.org/abs/2501.08617