Research map / Optimization / Optimizers
Textual Stochastic Gradient Descent: Discrete Optimization of External Memory for Reasoning Language Agents
Jian Li, Hua Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add TSGD experience library optimization | Grok-4.1 · MATH Level 5 overall | accuracy: 49.30 -> 68.0 | Improved |
| add initialization of experience library (ablation) | TSGD · MATH Precalculus L5 | Pass@1: 33.52 -> 55.67 | Improved |
| add optimization without dual verification (ablation) | TSGD (init + optimization) · MATH Precalculus L5 | Pass@1: 55.67 -> 47.21 | Hurt |
| add dual verification (ablation) | TSGD · MATH Precalculus L5 | Pass@1: 47.21 -> 53.61 | Improved |
| add periodic regularization under capacity budget (ablation) | TSGD · MATH Precalculus L5 | Pass@1: 53.61 -> 58.33 | Improved |
| add periodic regularization under capacity budget (ablation) | TSGD · MATH Precalculus L5 | library size (B): 204 -> 27 | Improved |
| add subject filtering to retrieval (ablation) | retrieval agent · MATH Precalculus | Mean@1: 49.60% -> 56.17% | Improved |
| retrieve high-relevance rather than low-relevance experiences (ablation) | TSGD · MATH Precalculus L5 | Pass@1: 49.80 -> 59.76 | Improved |
| retrieve low-relevance (Geometry) experiences (ablation) | TSGD · MATH Precalculus L5 | Pass@1: 33.35 -> 49.80 | Improved |
| increase retrieved experiences k beyond about 14 (ablation) | TSGD · MATH | Mean@32 | No clear effect |
| add global verification to reject candidate updates (ablation) | TSGD · optimization training | update acceptance: ≈28% rejected | Improved |
| add periodic regularization versus unregularized growth (ablation) | TSGD · training stability | stability of improvement | Improved |
Closest papers by meaning
- GReaTer: Gradients Over Reasoning Makes Smaller Language Models Strong Prompt Optimizers (ICLR 2025 · similarity 0.88)
- Reasoning via Test-Time Instance-Level Policy Gradient in Latent Space (NeurIPS 2026 · similarity 0.88)
- Optimizing Language Models for Inference Time Objectives using Reinforcement Learning (ICML 2025 · similarity 0.88)
- Thinking on the Fly: Test-Time Reasoning Enhancement via Latent Thought Policy Optimization (ICLR 2026 · similarity 0.88)
- $\nabla$-Reasoner: LLM Reasoning via Test-Time Gradient Descent in Textual Space (ICLR 2026 · similarity 0.88)
- MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
