Start / LessWrong (30+ Karma) / Linkpost research note our scheming precursor evals had limited predictive power for our in context scheming evals by marius hobbhahn

[Linkpost] “Research Note: Our scheming precursor evals had limited predictive power for our in-context scheming evals” by Marius Hobbhahn

3 min • 3 juli 2025

This is a link post.

Note: This is a research note, and the analysis is less rigorous than our standard for a published paper. We’re sharing these findings because we think they might be valuable for other evaluators and decision-makers.

Executive Summary

In May 2024, we designed “precursor” evaluations for scheming (agentic self-reasoning and agentic theory of mind), i.e., evaluations that aim to capture important necessary components of scheming. In December 2024, we published “in-context scheming” evaluations, i.e. evaluations that directly aim to measure scheming reasoning capabilities. We have easy, medium, and hard difficulty levels for all evals.
In this research note, we run some basic analysis on how predictive our precursor evaluations were of our scheming predictions to test the underlying hypothesis of whether the precursor evals would have “triggered” relevant scheming thresholds.
We run multiple pieces of analysis to test whether our precursor evaluations predict our scheming [...]

---

First published:
July 3rd, 2025

Source:
https://www.lesswrong.com/posts/9tqpPP4FwSnv9AWsi/research-note-our-scheming-precursor-evals-had-limited

Linkpost URL:
https://www.apolloresearch.ai/blog/research-note-our-scheming-precursor-evals-had-limited-predictive-power-for-our-in-context-scheming-evals

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Three data visualizations showing miss rates and false alarm rates analysis: two heatmaps and a scatter plot with performance thresholds.

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Senaste avsnitt

“against that one rationalist mashal about japanese fifth-columnists” by Fraser

13 juli | 6 min

“Surprises and learnings from almost two months of Leo Panickssery” by Nina Panickssery

13 juli | 12 min

“Vitalik’s Response to AI 2027” by Daniel Kokotajlo

12 juli | 24 min

“the jackpot age” by thiccythot

12 juli | 13 min

“Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity” by habryka

11 juli | 12 min

[Linkpost] “Research Note: Our scheming precursor evals had limited predictive power for our in-context scheming evals” by Marius Hobbhahn

Senaste avsnitt

“against that one rationalist mashal about japanese fifth-columnists” by Fraser

“Surprises and learnings from almost two months of Leo Panickssery” by Nina Panickssery

“Vitalik’s Response to AI 2027” by Daniel Kokotajlo

“the jackpot age” by thiccythot

“Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity” by habryka

[Linkpost] “Guide to Redwood’s writing” by Julian Stastny

“So You Think You’ve Awoken ChatGPT” by JustisMills

[Linkpost] “Open Global Investment as a Governance Model for AGI” by Nick Bostrom

“what makes Claude 3 Opus misaligned” by janus

“Lessons from the Iraq War about AI policy” by Buck

“Generalized Hangriness: A Standard Rationalist Stance Toward Emotions” by johnswentworth

“Evaluating and monitoring for AI scheming” by Vika, Scott Emmons, Erik Jenner, Mary Phuong, Lewis Ho, Rohin Shah

“White Box Control at UK AISI - Update on Sandbagging Investigations” by Joseph Bloom, Jordan Taylor, Connor Kissane, Sid Black, merizian, alexdzm, jacoba, Ben Millwood, Alan Cooney

“80,000 Hours is producing AI in Context — a new YouTube channel. Our first video, about the AI 2027 scenario, is up!” by chanamessinger

[Linkpost] “No, We’re Not Getting Meaningful Oversight of AI” by Davidmanheim

“What’s worse, spies or schemers?” by Buck, Julian Stastny

“No, Grok, No” by Zvi

“Applying right-wing frames to AGI (geo)politics” by Richard_Ngo

“A deep critique of AI 2027’s bad timeline models” by titotal

“Subway Particle Levels Aren’t That High” by jefftk

“An Opinionated Guide to Using Anki Correctly” by Luise

“Why Do Some Language Models Fake Alignment While Others Don’t?” by abhayesian, John Hughes, Alex Mallen, Jozdien, janus, Fabien Roger

“Balsa Update: Springtime in DC” by Zvi

[Linkpost] “A Theory of Structural Independence” by Matthias G. Mayer

“On Alpha School” by Zvi

“You Can’t Objectively Compare Seven Bees to One Human” by J Bostock

“Literature Review: Risks of MDMA” by Elizabeth

“45 - Samuel Albanie on DeepMind’s AGI Safety Approach” by DanielFilan

“On the functional self of LLMs” by eggsyntax

“Shutdown Resistance in Reasoning Models” by benwr, JeremySchlatter, Jeffrey Ladish

“The Cult of Pain” by Martin Sustrik

[Linkpost] “Claude is a Ravenclaw” by Adam Newgas

“‘Buckle up bucko, this ain’t over till it’s over.’” by Raemon

“How much novel security-critical infrastructure do you need during the singularity?” by Buck

“‘AI for societal uplift’ as a path to victory” by Raymond Douglas

“Two proposed projects on abstract analogies for scheming” by Julian Stastny

“Outlive: A Critical Review” by MichaelDickens

“Authors Have a Responsibility to Communicate Clearly” by TurnTrout

[Linkpost] “MIRI Newsletter #123” by Harlan, Rob Bensinger

[Linkpost] “Research Note: Our scheming precursor evals had limited predictive power for our in-context scheming evals” by Marius Hobbhahn

“Call for suggestions - AI safety course” by boazbarak

[Linkpost] “IABIED: Advertisement design competition” by yams

“Congress Asks Better Questions” by Zvi

“Curing PMS with Hair Loss Pills” by David Lorell

“AI Task Length Horizons in Offensive Cybersecurity” by Sean Peters

“Race and Gender Bias As An Example of Unfaithful Chain of Thought in the Wild” by Adam Karvonen, Sam Marks

“There are two fundamentally different constraints on schemers” by Buck

“‘What’s my goal?’” by Raemon

“A Simple Explanation of AGI Risk” by TurnTrout

“AI Moratorium Stripped From BBB” by Zvi

“Scientific Discovery in the Age of Artificial Intelligence” by Jessica Rumbelow

“SLT for AI Safety” by Jesse Hoogland

“The best simple argument for Pausing AI?” by Gary Marcus

“SAE on activation differences” by Santiago Aranguri, jacob_drori, Neel Nanda

“What We Learned Trying to Diff Base and Chat Models (And Why It Matters)” by Clément Dumas, Julian Minder, Neel Nanda

“If you want to be vegan but you worry about health effects of no meat, consider being vegan except for mussels/oysters” by KatWoods

[Linkpost] “Project Vend: Can Claude run a small shop?” by Gunnar_Zarncke

“Paradigms for computation” by Cole Wyeth

“life lessons from poker” by thiccythot

“Circuits in Superposition 2: Now with Less Wrong Math” by Linda Linsefors, Lucius Bushnaq

“I underestimated safety research speedups from safe AI” by Dan Braun

“Conciseness Manifesto” by Vasyl Dotsenko

“Support for bedrock liberal principles seems to be in pretty bad shape these days” by Max H

[Linkpost] “A Depressed Shrink Tries Shrooms” by AlphaAndOmega

[Linkpost] “[Paper] Stochastic Parameter Decomposition” by Lee Sharkey, Lucius Bushnaq, Dan Braun

“Childhood and Education #11: The Art of Learning” by Zvi

“Proposal for making credible commitments to AIs.” by Cleo Nardo

“Epoch: What is Epoch?” by Zach Stein-Perlman

“Recent and forecasted rates of software and hardware progress” by elifland

“Jankily controlling superintelligence” by ryan_greenblatt

“Help the AI 2027 team make an online AGI wargame” by Jonas V

“A Guide For LLM-Assisted Web Research” by nikos, dschwarz, Lawrence Phillips, FutureSearch

“If Not Now, When?” by Yair Halberstadt

“A case for courage, when speaking of AI danger” by So8res

“The Industrial Explosion” by rosehadshar, Tom Davidson

“Summary of John Halstead’s Book-Length Report on Existential Risks From Climate Change” by Bentham’s Bulldog

“Tech for Thinking” by sarahconstantin

“Lurking in the Noise” by J Bostock