AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Don't Walk the Line: Boundary Guidance for Filtered Generation

Sarah Ball, Andreas Haupt

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add Boundary Guidance fine-tuning (reward model + guard margin)Qwen2.5-0.5B · Filtered-generation benchmarkhelpfulness / harmfulness score: 2.13 -> 2.26 helpful; 0.27 -> 0.17 harmfulImproved
add Boundary Guidance fine-tuning (reward model + guard margin)Qwen2.5-14B · Filtered-generation benchmarkhelpfulness score: 2.77 -> 2.71No clear effect
add Boundary Guidance fine-tuning (reward model + guard margin)Qwen2.5-7B · Filtered-generation benchmarkharmfulness score: 0.19 -> 0.04Improved
add reward model to guard-only margin reward (ablation)Qwen2.5-0.5B · Filtered-generation benchmarkhelpfulness score: 1.40 -> 2.26Improved
add reward model to guard-only margin reward (ablation)Qwen2.5-7B · Filtered-generation benchmarkhelpfulness / harmfulness score: 2.75 -> 2.75 helpful; 0.03 -> 0.04 harmfulNo clear effect
use prompt-classifier-based sign (prompt-aware reward) instead of completion-based (ablation)Qwen2.5-14B · Filtered-generation benchmarkharmfulness score: 0.14 -> 0.27Hurt
use prompt-classifier-based sign (prompt-aware reward) instead of completion-based (ablation)Qwen2.5-7B · Filtered-generation benchmarkhelpfulness score: 2.72 -> 2.48Hurt
add HotpotQA long-context prompts to training mix (ablation)Guard-only Qwen2.5-7B · Filtered-generation benchmarkhelpfulness score: +0.03 -> +0.07Improved
add HotpotQA long-context prompts to training mix (ablation)Guard-only Qwen2.5-7B · Filtered-generation benchmarkharmfulness score: 0.03 -> 0.05Hurt
use ShieldGemma-2B instead of Llama-Guard-2-8B as classifier (ablation)Guard-only Qwen2.5-7B · Filtered-generation benchmarkhelpfulness score: 2.75 -> 2.10Hurt
use Granite-Guardian-125m instead of Llama-Guard-2-8B as classifier (ablation)Guard-only Qwen2.5-7B · Filtered-generation benchmarkharmfulness score: 0.03 -> 0.14 FT harmful; helpful 2.75 -> 2.71Mixed
shift guard:reward weighting from 1:1 to 2:1 (ablation)Boundary Guidance Qwen2.5-7B · Filtered-generation benchmarkhelpfulness / harmfulness score: 2.75 -> 2.74 helpful; 0.03 -> 0.04 harmfulNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy