Research map / Large language models / Retrieval and knowledge
When Silence Is Golden: Can LLMs Learn to Abstain in Temporal QA and Beyond?
Xinyu Zhou, Chang Jin, Carsten Eickhoff, Zhijiang Guo, Seyed Ali Bahrainian
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 CoT-SFT cold start then GRPO RL with abstention-aware reward | Qwen2.5-1.5B-Instruct · TimeQA-Hard | Exact Match (%): 29.95 -> 35.75 (vs GPT-4o) | Improved |
| use time-relevant sub-context instead of original context (ablation) | GPT-4o · TimeQA-Hard | Exact Match (%): 29.95 -> 34.12 | Improved |
| use time-relevant sub-context instead of original context (ablation) | GPT-3.5 · TimeQA-Hard | Exact Match (%): 21.09 -> 18.52 | Hurt |
| provide knowledge graph facts instead of the original context (ablation) | GPT-4o · TimeQA-Easy | Exact Match (%): 39.95 -> 30.10 | Hurt |
| extract KGs from whole context instead of chunked context (ablation) | GPT-4o · TimeQA-Easy | Exact Match (%): 24.66 -> 30.10 | Improved |
| supervised fine-tuning on context | Qwen2.5-1.5B-Instruct · TimeQA-Easy | Exact Match (%): 8.06 -> 1.81 | Hurt |
| add original context to question-only prompt (ablation) | Qwen2.5-7B-Instruct · TimeQA-Hard | Exact Match (%): 11.69 -> 9.00 | Hurt |
| use ROUGE+EM answer reward instead of ROUGE only (ablation) | RL, Qwen2.5-1.5B-Instruct · TimeQA-Hard | Exact Match (%): 26.67 -> 35.75 | Improved |
| add BERTScore to the ROUGE+EM reward (ablation) | RL, Qwen2.5-1.5B-Instruct · TimeQA-Easy vs Hard | Exact Match (%): Easy 43.41 -> 33.71; Hard 35.75 -> 37.62 | Mixed |
| raise KL coefficient from 0.01 to 0.04 (ablation) | RL, Qwen2.5-1.5B-Instruct · TimeQA-Easy | Exact Match (%): 42.74 -> 41.56 | Hurt |
| increase unanswerable share in RL training data to 50% (ablation) | RL, Qwen2.5-1.5B-Instruct · TimeQA-Easy | Exact Match (%): 43.41 -> 11.24 | Hurt |
| use LoRA instead of full-parameter SFT (ablation) | Qwen2.5-1.5B-Instruct · TimeQA-Easy | False positives (count): 309 -> 437 | Hurt |
Closest papers by meaning
- AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions (NeurIPS 2025 · similarity 0.93)
- Knowing When to Quit: A Principled Framework for Dynamic Abstention in LLM Reasoning (ICML 2026 · similarity 0.91)
- Test of Time: A Benchmark for Evaluating LLMs on Temporal Reasoning (ICLR 2025 · similarity 0.90)
- Thinker: Learning to Think Fast and Slow (NeurIPS 2025 · similarity 0.89)
- LoongRL: Reinforcement Learning for Advanced Reasoning over Long Contexts (ICLR 2026 · similarity 0.89)
- Can Speech LLMs Think while Listening? (ICLR 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.
