We're sunsetting PodQuest on 2025-07-28. Thank you for your support!

“Inference-Time-Compute: More Faithful? A Research Note” by James Chua, Owain_Evans

2025/1/15

LessWrong (30+ Karma)

AI Chapters

Chapters

Shownotes Transcript

No transcript made for this episode yet, you may request it for free.

“Inference-Time-Compute: More Faithful? A Research Note” by James Chua, Owain_Evans

LessWrong (30+ Karma)

Abstract: What's the Core of This Research?

Introduction: Setting the Stage

Setup and Results of Cues: How Do Models Respond to Influential Cues?

Cue: Professor's Opinion: Does It Matter?

Cue: Few-Shot with Black Square: A Unique Test

Other Cues: Exploring Further Influences

Discussion: What Do the Results Mean?

Improving Non-ITC Articulation: Can We Bridge the Gap?

Advantage of ITC Models in Articulation: Why Are They Better?

Length of CoTs Across Models: Does Size Matter?

False Positives: What Are the Pitfalls?

Different Articulation Rates Across Cues: Why the Variance?

Training Data Contamination: A Hidden Issue?

Limitations: What Are the Challenges?

Lack of ITC Models to Evaluate: A Limiting Factor?

Limited Cues Studied: Is the Scope Too Narrow?

Subjectivity of Judge Model: How Reliable?

Acknowledgments: Who Contributed?

Links: Where to Find More?

Shownotes Transcript

“Inference-Time-Compute: More Faithful? A Research Note” by James Chua, Owain_Evans 26:09 Share

LessWrong (30+ Karma)

Abstract: What's the Core of This Research?

Introduction: Setting the Stage

Setup and Results of Cues: How Do Models Respond to Influential Cues?

Cue: Professor's Opinion: Does It Matter?

Cue: Few-Shot with Black Square: A Unique Test

Other Cues: Exploring Further Influences

Discussion: What Do the Results Mean?

Improving Non-ITC Articulation: Can We Bridge the Gap?

Advantage of ITC Models in Articulation: Why Are They Better?

Length of CoTs Across Models: Does Size Matter?

False Positives: What Are the Pitfalls?

Different Articulation Rates Across Cues: Why the Variance?

Training Data Contamination: A Hidden Issue?

Limitations: What Are the Challenges?

Lack of ITC Models to Evaluate: A Limiting Factor?

Limited Cues Studied: Is the Scope Too Narrow?

Subjectivity of Judge Model: How Reliable?

Acknowledgments: Who Contributed?

Links: Where to Find More?

Shownotes Transcript

“Inference-Time-Compute: More Faithful? A Research Note” by James Chua, Owain_Evans