Record · analyzed
TextGrad: Automatic "Differentiation" via Text
Uses written feedback from a model as a "gradient" to edit prompts, code or answers: GPQA 51% → 55% for GPT-4o.
1
Listed
Added to the list. How entries were checked when they were listed is in the journal.
2
Collected
| File | Source | SHA-256 |
|---|---|---|
| paper PDF2406.07496.pdf | arxiv.org/pdf/2406.07496 | 9cbfd5c78ad69e2a8363e76d6774f3e6b5e68f46b409d8b560e98276a985b428 |
To check a file, download it from its source and compare its SHA-256 with the one above. Collected copies are kept in the editor's RSI system; manifest: design-life/literature/rsi-workflow-improvement/sources.json.
3
Analyzed
Quoted verbatim from design-life/literature/rsi-workflow-improvement/comparison.md (2026-10-07). Parts about applying the paper to the editor's private work are left out.
From “Evidence for improving reusable agent workflows” (2026-10-07)
Inspected edition: TextGrad, v1, 11 June 2024; PDF pp. 3–8, 17
Mechanism and reported evidence: Routes natural-language feedback through component dependencies. On 39 LeetCode-Hard problems with GPT-4o, Table 1 reports completion 0.36 versus Reflexion 0.31, both with five iterations; averaged over five seeds. Completion means passing platform tests.
4
Replicated
Not done yet. Not attempted yet.
5
Audited
Not done yet. Not attempted yet.
The editor's view
None recorded yet.