Research map / Architectures and efficiency / RNNs and normalization
Detecting the Semantic Fixed Point: A Geometric Framework for Efficient Inference
Jiawei Gu, Ziyue Qiao, Xiao Luo
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 |
|---|---|---|---|
| use normalized step size together with directional stability instead of step size only (ablation) | GCEE · TriviaQA | retention: 89.1% -> 98.4% | Improved |
| add step size to direction-only criterion (ablation) | GCEE · TriviaQA | retention: 84.0% -> 98.4% | Improved |
| add normalization of step size (ablation) | GCEE · TriviaQA | retention: 80.8% -> 98.4% | Improved |
| use step-size plus direction criterion instead of hidden-state cosine similarity (ablation) | GCEE · TriviaQA | EM: 54.8 -> 63.5 | Improved |
| add safeguards (warmup l_min=4, K=2 consecutive layers) (ablation) | GCEE · TriviaQA | retention: 74.7% -> 98.4% | Improved |
| raise safeguards to l_min=6, K=3 (ablation) | GCEE · TriviaQA | retention: 98.4% -> 98.9% | No clear effect |
| monitor last token instead of mean pooling (ablation) | GCEE · TriviaQA | EM: 59.2 -> 63.5 | Improved |
| monitor last token instead of first token (ablation) | GCEE · TriviaQA | EM: 51.5 -> 63.5 | Improved |
| vary thresholds within tau_s 0.008-0.015 and tau_c 0.88-0.92 (ablation) | GCEE · TriviaQA | EM: EM varies by less than 1 point | No clear effect |
| use geometric criterion instead of entropy exit | GCEE · TriviaQA (LLaMA-2-7B) | EM: 58.9 -> 63.5 | Improved |
| use adaptive GCEE depth instead of static truncation at matched depth | GCEE · TriviaQA | accuracy: +8.3 | Improved |
| use GCEE instead of entropy exit for wall-clock speed | GCEE · TriviaQA | E2E speedup: 1.14x -> 1.45x | Improved |
Closest papers by meaning
- The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations (NeurIPS 2026 · similarity 0.87)
- Search for Truth from Reasoning: A Dynamic Representation Editing Framework for Steering LLM Trajectories (ICML 2026 · similarity 0.87)
- The Tell-Tale Norm: $\ell_2$ Magnitude as a Signal for Reasoning Dynamics in Large Language Models (ICML 2026 · similarity 0.87)
- Transformers Represent Belief State Geometry in their Residual Stream (NeurIPS 2024 · similarity 0.87)
- Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language Models (ICML 2026 · similarity 0.87)
- Boosting Inference with Guided Reasoning: Stochastic Exploration for Recursive Models (NeurIPS 2026 · similarity 0.86)
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.
