Research map / Large language models / Instruction and fine-tuning
Don't Walk the Line: Boundary Guidance for Filtered Generation
Sarah Ball, Andreas Haupt
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 Boundary Guidance fine-tuning (reward model + guard margin) | Qwen2.5-0.5B · Filtered-generation benchmark | helpfulness / harmfulness score: 2.13 -> 2.26 helpful; 0.27 -> 0.17 harmful | Improved |
| add Boundary Guidance fine-tuning (reward model + guard margin) | Qwen2.5-14B · Filtered-generation benchmark | helpfulness score: 2.77 -> 2.71 | No clear effect |
| add Boundary Guidance fine-tuning (reward model + guard margin) | Qwen2.5-7B · Filtered-generation benchmark | harmfulness score: 0.19 -> 0.04 | Improved |
| add reward model to guard-only margin reward (ablation) | Qwen2.5-0.5B · Filtered-generation benchmark | helpfulness score: 1.40 -> 2.26 | Improved |
| add reward model to guard-only margin reward (ablation) | Qwen2.5-7B · Filtered-generation benchmark | helpfulness / harmfulness score: 2.75 -> 2.75 helpful; 0.03 -> 0.04 harmful | No clear effect |
| use prompt-classifier-based sign (prompt-aware reward) instead of completion-based (ablation) | Qwen2.5-14B · Filtered-generation benchmark | harmfulness score: 0.14 -> 0.27 | Hurt |
| use prompt-classifier-based sign (prompt-aware reward) instead of completion-based (ablation) | Qwen2.5-7B · Filtered-generation benchmark | helpfulness score: 2.72 -> 2.48 | Hurt |
| add HotpotQA long-context prompts to training mix (ablation) | Guard-only Qwen2.5-7B · Filtered-generation benchmark | helpfulness score: +0.03 -> +0.07 | Improved |
| add HotpotQA long-context prompts to training mix (ablation) | Guard-only Qwen2.5-7B · Filtered-generation benchmark | harmfulness score: 0.03 -> 0.05 | Hurt |
| use ShieldGemma-2B instead of Llama-Guard-2-8B as classifier (ablation) | Guard-only Qwen2.5-7B · Filtered-generation benchmark | helpfulness score: 2.75 -> 2.10 | Hurt |
| use Granite-Guardian-125m instead of Llama-Guard-2-8B as classifier (ablation) | Guard-only Qwen2.5-7B · Filtered-generation benchmark | harmfulness score: 0.03 -> 0.14 FT harmful; helpful 2.75 -> 2.71 | Mixed |
| shift guard:reward weighting from 1:1 to 2:1 (ablation) | Boundary Guidance Qwen2.5-7B · Filtered-generation benchmark | helpfulness / harmfulness score: 2.75 -> 2.74 helpful; 0.03 -> 0.04 harmful | No clear effect |
Closest papers by meaning
- Adaptive Classifier-Free Guidance via Dynamic Low-Confidence Masking (NeurIPS 2025 · similarity 0.87)
- Diffusion Guidance Is a Controllable Policy Improvement Operator (NeurIPS 2026 · similarity 0.85)
- Beyond the Dirac Delta: Mitigating Diversity Collapse in Reinforcement Fine-Tuning for Image Generation (NeurIPS 2026 · similarity 0.85)
- Visual Generation Without Guidance (ICML 2025 · similarity 0.84)
- Don't be so Negative! Score-based Generative Modeling with Oracle-assisted Guidance (ICML 2024 · similarity 0.84)
- Policy-shaped prediction: avoiding distractions in model-based reinforcement learning (NeurIPS 2024 · similarity 0.84)
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.
