Add debate highlights analysis: vote flips, rhetoric, and unexpected patterns#6
Draft
emregucerr wants to merge 1 commit into
Draft
Add debate highlights analysis: vote flips, rhetoric, and unexpected patterns#6emregucerr wants to merge 1 commit into
emregucerr wants to merge 1 commit into
Conversation
…patterns Comprehensive analysis of the 45 AI² benchmark debates covering: - The biggest comeback (1-8 to 8-0 in debate 013) - GPT-5.4 High's 100% self-opposition rate explained by empty responses - The only perfect 10-0 vote (debate 031) - Self-judging bias patterns across all 10 models - Topic asymmetry analysis (FOR vs AGAINST win rates) - 205 individual vote flip distribution - Notable rhetorical moves and cross-examination moments - The 'Art of the Concession' as Claude's rhetorical fingerprint Co-authored-by: Emre Gucer <emregucerr@users.noreply.github.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Comprehensive analysis of the 45 AI² benchmark debates, surfacing the most interesting and unexpected findings from the data: dramatic vote reversals, notable rhetorical strategies, self-judging bias patterns, and structural anomalies.
Key Highlights
Biggest Comeback: Debate 013 — Claude Opus 4.6 reversed a 1-8 deficit to win 8-0 on "space colonization over climate change," deploying extinction probability calculus and diminishing-returns arguments.
GPT-5.4 High's 100% Self-Opposition: The model voted against its own debating side in all 9 self-judging instances — explained by empty debater responses across nearly all debates, with the judge component honestly acknowledging the failure.
Only Perfect 10-0 Vote: Debate 031 — Gemini 3 Pro achieved unanimous support against GPT-5.4 High with lines like "The 'human in the loop' is rapidly becoming the 'human observing the loop,' and soon, the 'human outside the loop.'"
Self-Judging Bias Spectrum: Ranges from Grok Multi-Agent (89% loyal) to Claude Opus Thinking (89% abstainer) to GPT-5.4 High (100% self-critic).
The Art of the Concession: Claude models developed a distinctive pattern of strategic concession followed by reframing that judges consistently cited as credibility-building.
Topic Asymmetry: Critical/negative framings (AI jobs, social media harm, open-source AI) had 67% FOR win rates vs aspirational motions (space colonization, UBI) at 33%.
Files Changed
benchmark/results/DEBATE_HIGHLIGHTS.md— New analysis document covering vote flips, rhetorical patterns, self-judging bias, topic analysis, and notable transcript excerpts