Stop Polishing, Start Switching
By Lovro Lucic ·
The "What You Think Works" Problem · 1 of 1
Polishing doesn't work past a point. The output hits a ceiling and more iteration produces diminishing returns. The first response isn't the best but the fifth response isn't much better than the third.
The ceiling is per generation mode. Not per model. Not per session. Per mode.
When you ask the model to analyze, it generates in analytical mode. Each iteration within analytical mode improves the analysis marginally. The structure tightens. The language gets cleaner. The coverage gets more complete. But the analytical depth doesn't increase because the model is iterating within the same semantic region. It's polishing, not discovering.
Switch to a different mode: ask for critique, ask for the opposite argument, ask it to find what's wrong: and the output jumps to a different region. Not because the model got smarter. Because it accessed a different part of representation space. The ceiling in analytical mode doesn't apply in critical mode. Each mode has its own ceiling.
The practical version: when output stops improving, don't ask for another iteration in the same mode. Switch modes. "Now critique what you just produced." "What's wrong with this analysis?" "Argue the opposite." "What did this miss?" The mode switch accesses territory that iteration within a single mode can't reach.
This connects to why self-critique circles instead of improving. When you ask the model to critique its own output in the same generation context, it's often still in the original mode. The "critique" activates a critique-flavored version of the same semantic neighborhood, not a genuinely different critical perspective. A fresh prompt with an explicitly different mode produces better critique than "now review what you just wrote."
Three mode switches that break ceilings in practice (the first is tested at scale, the other two are consistent observations). Analytical → critical: "What's wrong with this?" Generative → evaluative: "Score each option against these criteria." Convergent → divergent: "What's a completely different approach?" Each switch accesses a different region. The output after the switch contains information that wasn't accessible from the previous mode.
A non-adversarial mode switch test resolved the one confound in the original experiment. The original switched analytical to critical, which introduces contrary vocabulary by design. The follow-up switched analytical to evaluative (scoring options against criteria). No adversarial content. The evaluative switch produced more novelty than continued iteration on all 5 topics tested, and exceeded fresh context on all 5. The novelty boost comes from the mode switch itself, not from the adversarial vocabulary the critical mode introduces.
You polished past the point of return. The third iteration was marginal. The fifth was wasted. But iteration feels like progress. The ceiling is per mode, not per session. You stayed in the same mode because switching feels like giving up.
Narrowing Test (#7). In your last AI session where you iterated: did you change your fundamental approach at any point, or did every iteration refine the same direction? Count: how many iterations were refinement vs how many were genuine reframes.
Test this yourself
Next time iteration five looks barely better than iteration three, stop polishing. Ask: "What's wrong with this?" One mode switch, then compare.
What survived testing
- Iteration within a mode produces diminishing returnsCopy link
- Self-critique in the same context circlesCopy link
- Mode switch produces more novel vocabulary than continued iteration, winning on all topics tested across two generators. Density-normalized to control for length.Copy link
- Mode switch exceeds fresh context. The conversation provides material to push against, producing more novelty than starting fresh.Copy link
- Cross-generator: both models show the effect. Larger on one.Copy link
- Non-adversarial switch (analytical to evaluative) confirms mechanism: the novelty boost comes from the mode switch itself, not adversarial vocabulary.Copy link
What didn't survive
Honest limits
- Mode switch tested with analytical→critical (adversarial) only. The adversarial prompt introduces contrary vocabulary by design. The novelty could partly come from the content shift, not purely from mode switching. The "exceeds fresh context" finding argues against pure content shift (fresh context has no conversation to argue against), but the confound isn't fully resolved.Copy link
- Only one mode-switch direction tested (analytical→critical). Generative→evaluative and convergent→divergent remain observational.Copy link
- "Per generation mode" is an explanatory model. The actual representation space dynamics are more complex.Copy link
- March 2026 models. Future models may have less pronounced mode boundaries.Copy link
Explore other threads
The Fabrication Problem
4 findingsMost AI numbers are fabricated. Source material fixes it. Self-checking fails. Trust signals are backwards.
The Evaluation Problem
2 findingsJudgment goes quiet. You can't see the gaps. Satisfaction is the trap. Stronger evaluators discriminate less.
The "It Depends" Problem
3 findingsSame instruction, opposite results. Specificity is the lever. Context redirects, not informs. The measurement itself was wrong.
New findings when they land.
No spam. Just what held up.