Research map / Large language models / Language, safety and interpretability
CGRiC: Compositional Risk Certification for Structured LLM Outputs
Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma
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 |
|---|---|---|---|
| upgrade to retrieval+NLI verifier (ablation) | Atomic EDFL · HotpotQA-Cite | abstention rate: .42 -> .39 | Improved |
| add per-claim granularity with max-risk composition (ablation) | stronger-verifier atomic baseline · HotpotQA-Cite | abstention rate: .39 -> .37 | Improved |
| add union-bound composition and localized repair (ablation) | per-claim retrieval+NLI baseline · HotpotQA-Cite | abstention rate: .37 -> .29 | Improved |
| replace atomic EDFL with CGRiC (union) | EDFL · HotpotQA | abstention rate: .42 -> .29 | Improved |
| use graph-structured bound instead of union bound (ablation) | CGRiC · HotpotQA | abstention rate: .29 -> .26 | Improved |
| add robustness extension over skeleton family (ablation) | CGRiC (union) · HotpotQA | abstention rate: .29 -> .33 | Hurt |
| use CGRiC instead of EDFL under adversarial falsehoods | EDFL · HotpotQA-Cite adversarial falsehoods | certified-failure rate: .18 -> .083 | Improved |
| distribution shift from in-distribution to adversarial falsehoods | CGRiC · HotpotQA-Cite calibration | certified-failure rate: .050 -> .083 | Hurt |
| 20% drop noise in claim extraction (ablation) | CGRiC · extraction noise | certified-failure rate: 0.05 -> 0.08 | Hurt |
| use lazy CGRiC fast path (ablation) | CGRiC · computational overhead | latency overhead: +35-47% -> +18-26% | Improved |
| add localized repair instead of full regeneration (ablation) | CGRiC · latency overhead | latency overhead: +85% -> +42% | Improved |
Closest papers by meaning
- When Should LLMs Be Less Specific? Selective Abstraction for Reliable Long-Form Text Generation (NeurIPS 2026 · similarity 0.86)
- CITE: Anytime Valid Statistical Inference in LLM Self-Consistency (NeurIPS 2026 · similarity 0.83)
- Graph-based Uncertainty Metrics for Long-form Language Model Generations (NeurIPS 2024 · similarity 0.83)
- Grammars of Formal Uncertainty: When to Trust LLMs in Automated Reasoning Tasks (NeurIPS 2025 · similarity 0.82)
- From Judgment to Interference: Early Stopping LLM Harmful Outputs via Streaming Content Monitoring (NeurIPS 2025 · similarity 0.82)
- SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks (NeurIPS 2025 · similarity 0.82)
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.
