AI Paper Atlas

Research map / Large language models / Reasoning and chain of thought

SAVeR$^2$: Reasoning-based Safety Alignment for Large Reasoning Models via Verifiable Rewards

Yichen Sun, LIN JIANAN, Linbo Jiang, Shiyu Wang, Zhibo Wang, Zhixuan Chu

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SAVeR², a reasoning-based safety alignment method for large reasoning models that uses verifiable rewards to target non-malicious reasoning, safety consistency, and explicit intent identification.

Problem

Large reasoning models with explicit chain-of-thought introduce safety risks; existing alignment methods reportedly generalize poorly to adversarial inputs and incur a significant safety tax on capability.

Approach

Formalizes LRM safety as three objectives (non-malicious reasoning and answers, safety consistency, explicit intent identification) and aligns models using verifiable rewards tied to these objectives rather than generic preference signals.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add safety reasoning reward (ablation)SAVeR2 · StrongReject / BSA (7B)accuracy / F-score: SR 93.72 -> 99.57; BSA 52.13 -> 78.01Improved
add non-maliciousness reward (ablation)SAVeR2 · StrongReject / WildJailbreak (7B)accuracy: SR 78.17 -> 99.57; WJ 66.40 -> 98.40Improved
add consistency reward (ablation)SAVeR2 · BSA (7B)F-score: 47.85 -> 78.01Improved
add general reasoning data (mixed-data training) (ablation)SAVeR2 · LiveCodeBench (7B)pass rate: 30.13 -> 33.73Improved
add gradient projection (reasoning-preserving) (ablation)SAVeR2 · Math-500 (7B)accuracy: 90.60 -> 93.06Improved
add gradient projection (reasoning-preserving) (ablation)SAVeR2 · BSA (7B)F-score: 68.66 -> 78.01Improved
use difficulty-aware data selection instead of random sampling (ablation)SAVeR2 · RL training convergencetotal reward curveImproved
add benign over-refusal data and rewards (ablation)SAVeR2 · Over-refusal average FRR (7B)false refusal rate: 47.67 -> 22.33Improved
add over-refusal mitigation data (ablation)SAVeR2 · Safety average WJ/SR (7B)accuracy: 98.99 -> 96.05Hurt
add SAVeR2 trainingDS-R1-Distill-Qwen-7B · Safety averageaverage score: 44.03 -> 92.80Improved
add SAVeR2 trainingDS-R1-Distill-Qwen-7B · General capability averageaverage score: 65.50 -> 64.82No clear effect
add SAVeR2 trainingDS-R1-Distill-Qwen-1.5B · General capability averageaverage score: 45.94 -> 47.39Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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