Effective Feedback Compute
New research introduces Effective Feedback Compute (EFC), challenging traditional metrics by showing that AI performance depends more on how feedback is used than on raw compute power. EFC predicts failure rates with R² of 0.94, far outperforming token counts, and boosts success rates from 0.27 to 0.90 when feedback quality improves.
Article intelligence
Key points
- EFC measures the efficiency of feedback use, outperforming raw compute metrics in predicting AI failure rates
- Oracle-EFC achieved R²=0.94 in controlled tests, compared to 0.33 for raw token counts
- Improving feedback quality increased success rates from 0.27 to 0.90
- The study suggests a shift toward feedback-first AI development
Why it matters
This matters because EFC measures the efficiency of feedback use, outperforming raw compute metrics in predicting AI failure rates.
Technical impact
May affect model selection, inference cost, product capability, and evaluation benchmarks.
Effective Feedback Compute: The Real Game Changer in AI Performance
By Callum BryceMay 29, 2026
Share:
New research suggests that in AI, it's not about how much compute you use, but how smartly you use it. Effective Feedback Compute (EFC) is the buzzword that's reshaping AI scaling.
Forget the raw compute numbers. The future of AI performance lies in how efficiently feedback is used. That's the message from a groundbreaking study introducingEffective Feedback Compute(EFC). It challenges the old metrics of tokens and tool calls, making a case for feedback that's not just frequent, but genuinely informative and retained.
Why EFC Matters
Here's the deal: AI systems aren't just about cranking up the compute power. They're about what you do with the feedback you get. The study shows that EFC predicts failure rates far better than traditional methods. In a controlled environment, Oracle-EFC scored a staggering $R^2=0.94$, leaving raw token counts in the dust with a measly $0.33$.
This isn't just academic fluff. When feedback quality improved, success rates jumped from $0.27$ to $0.90$. So, why are we still obsessing over raw compute budgets? This is a call for a smarter approach.
The Numbers Don't Lie
Let's talk numbers. In mixed real trace tests, NRS-EFC/$D_{\mathrm{task}}$ achieved a $R^2=0.92$. Compare that to the near-zero impact of raw compute. And it's not just one-off success, EFC remained the top predictor in holdout test sets, scoring $R^2=0.85$.
What does this mean? It's simple. The AI race isn't just about who can spend more on compute. It's about who can spend wisely. And just like that, the leaderboard shifts.
The Bigger Picture
Isn't it wild that we've been measuring AI progress by sheer computational muscle? This study flips that narrative, suggesting a new way forward. It's not just a tweak in metrics. It's a potential shift in how AI development is approached.
So, what's next? Are AI developers ready to embrace a feedback-first approach? Because if the numbers are anything to go by, they should be. This changes the landscape.
Share this article:
Get AI news in your inbox
Daily digest of what matters in AI.
Key Terms Explained
Compute
The processing power needed to train and run AI models.
Token
The basic unit of text that language models work with.