Research map / Large language models / Retrieval and knowledge
RECKONING: Reasoning through Dynamic Knowledge Encoding
Zeming Chen, Gail Weiss, Eric Mitchell, Asli Celikyilmaz, Antoine Bosselut
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 |
|---|---|---|---|
| Encode contextual knowledge in model parameters using gradient updates, then answer from the updated model | RECKONING · ProofWriter and CLUTRR-SG | label accuracy: 99.4 -> 99.5; 99.2 -> 99.7; 99.6 -> 99.8; 98.1 -> 98.3; 96.9 -> 97.6; 90.3 -> 94.8 | Improved |
| Add a knowledge-recovery CLM objective to question-answering training (ablation) | RECKONING · ProofWriter and CLUTRR-SG | label accuracy: 99.4 -> 99.5; 98.3 -> 99.5; 98.3 -> 99.7; 99.1 -> 99.8; 96.0 -> 98.3; 90.2 -> 94.8 | Improved |
| Add a knowledge-recovery objective to RECKONING training (ablation) | RECKONING · ProofWriter with all distractors | label accuracy: ~18.2% | Improved |
| Replace dynamic per-step, per-layer learning rates with one static shared learning rate (ablation) | RECKONING · CLUTRR-SG | label accuracy: average drop of 34.2%; 45.5% drop for 4-hop; 39.5% drop for 6-hop | Hurt |
| Increase inner-loop knowledge-encoding steps from one to six (ablation) | RECKONING · CLUTRR-SG | label accuracy: 5.9% for 2-hop; 42.3% for 4-hop; 34.7% for 6-hop | Improved |
| Add all distractor facts to the knowledge context (ablation) | FT-ICR · ProofWriter | label accuracy: 28.6% | Hurt |
| Add all distractor facts to the knowledge context (ablation) | RECKONING · ProofWriter | label accuracy: 17.1% | Hurt |
| Use RECKONING rather than FT-ICR when all distractors are present | GPT-2-small · ProofWriter with distractors | average label accuracy: 70.9% -> 82.5% | Improved |
| Use RECKONING rather than in-context reasoning | GPT-2-XL with LoRA · ProofWriter 5-hop with all distractors | accuracy: 65% -> 70.2% | Improved |
| Encode shared knowledge once and reuse updated parameters for multiple questions | RECKONING 1-step · Synthetic random-letter task, 18 questions | wall-clock time (s): 2.0436 -> 0.6228 | Improved |
| Encode shared knowledge once and reuse updated parameters for multiple questions | RECKONING 4-step · Synthetic random-letter task, 18 questions | wall-clock time (s): 2.0436 -> 1.4839 | Improved |
| Use RECKONING rather than GPT-2 in-context reasoning | FOLIO · FOLIO | accuracy: 53.3 -> 54.9 | Improved |
Closest papers by meaning
- Reasoning with Latent Thoughts: On the Power of Looped Transformers (ICLR 2025 · similarity 0.89)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.88)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.88)
- LoongRL: Reinforcement Learning for Advanced Reasoning over Long Contexts (ICLR 2026 · similarity 0.88)
- Thinker: Learning to Think Fast and Slow (NeurIPS 2025 · similarity 0.88)
- Longer Context, Deeper Thinking: Uncovering the Role of Long-Context Ability in Reasoning (NeurIPS 2025 · 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.
