Enable more precise LLM error correction by having users directly edit the model's previous response, conditioning the next response on this corrected version. This approach reduces token overhead by 79% compared to traditional separate-feedback methods while fixing more errors in complex reasoning tasks.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Enable more precise LLM error correction by having users directly edit the model's previous response, conditioning the next response on this corrected version. This approach reduces token overhead by 79% compared to traditional separate-feedback methods while fixing more errors in complex reasoning tasks.
In-Place Feedback: Direct Response Editing for Precise Guidance
Traditional multi-turn feedback works by issuing separate corrections: "Your answer has an error on step 3. Here's the correct step." The model then tries to apply this external feedback to its previous response. The problem is that external feedback is ambiguous—the model often fails to apply corrections precisely, leaving errors uncorrected or introducing new mistakes.
In-Place Feedback inverts this: users directly edit the model's response text, marking corrections inline. The model then conditions on this edited version, receiving explicit guidance on where and how to fix errors. This dramatically improves correction accuracy while reducing token overhead.
"""
Identify and mark errors in response.
Args:
response: Model's response text
problem: Original problem statement
ground_truth: Correct answer or reference
Returns:
annotations: List of errors to correct
edited_response: Response with corrections marked
"""
# Detect errors using error detector (could be learned model, rule-based, etc.)
"""
Format the edited response as input for the next generation step.
Args:
edited_response: Fully corrected response
annotations: List of corrections applied
Returns:
formatted_input: Text showing what was corrected
"""
# Format: show original | corrected pairs for clarity
Fully automatic workflows (human editing not available)
Tasks where 79% token savings is negligible (compute-rich setting)
Token efficiency comparison:
Approach
Turns
Avg Tokens per Turn
Total Tokens
Single-turn baseline
1
150
150
Traditional feedback
3
180
540
In-place feedback
3
95
285
Savings
-
-47%
-47%
Accuracy improvements on reasoning tasks:
Task
Baseline
With In-Place Feedback
Gain
Math reasoning
52%
61%
+9%
Multi-step logic
48%
59%
+11%
Code debugging
55%
66%
+11%
Key hyperparameters:
Parameter
Default
Tuning Notes
max_turns
3
More iterations = more fixes but diminishing returns
min_confidence
0.7
Lower = fix more errors but risk false corrections
error_detection
automatic
Manual feedback also supported
Common pitfalls:
Cascade errors: If first correction is wrong, second round amplifies it. Validate error detection accuracy >85%.
Over-correction: Model becomes "correction-seeking" (generates intentionally flawed responses to trigger feedback). Use entropy regularization to prevent.
Feedback clarity: Formatted edits must be unambiguous. Test on 20 examples to ensure model correctly parses corrections.
Convergence failure: Sometimes iterations increase error rate. Cap iterations and use best-seen response if later responses worse.
Integration checklist:
Implement error detector; validate on 100 examples (accuracy >85%)
Test formatted feedback on 20 examples; verify model applies corrections
Run single-turn vs. in-place feedback comparison on test set
Measure token overhead: should be ~50% less than traditional feedback
Monitor accuracy improvement per turn (should be monotonic or plateau)
Evaluate on multi-turn reasoning tasks (math, logic, code)