---
title: "DexDICE: Reconciling Safety and... | SI Paper Desk"
url: https://thesimind.si/papers/2611.09433/
description: "DexDICE learns from two kinds of demonstrations, safe but weaker ones and strong but risky ones, to build agents that score well while staying inside safe..."
---

Number 8 of 12 · November 9, 2026

# DexDICE: Reconciling Safety and Performance via Dual-Expert Offline Imitation Learning

DexDICE learns from two kinds of demonstrations, safe but weaker ones and strong but risky ones, to build agents that score well while staying inside safe regions.

Seokin Seo, Sung Kuk Shyn, Minji Seo and 4 more · Submitted November 6, 2026 · arXiv 2611.09433v1 · cs.LG

[Read the paper on arXiv](https://arxiv.org/abs/2611.09433) · [The PDF](https://arxiv.org/pdf/2611.09433) · [The day's 12](https://thesimind.si/papers/day/2026-11-09/)

## What they found

The authors study offline imitation learning, where an agent learns only from recorded demonstrations. They propose "dual-expert" learning with two data sources: safety-compliant demonstrations that may be suboptimal, and performance-oriented demonstrations that may break safety guidelines. Their method, DexDICE, combines DICE-style offline learning (which corrects for differences in how often states are visited) with safe support constraints. These constraints keep the agent inside regions the safe data covers. They prove that the ideal solution, if the constraint is met exactly, has zero unsafe-region visitation with high probability. They test it on a re-curated DSRL benchmark and on a real mobile-robot navigation task trained from human teleoperation demonstrations. The abstract says it substantially outperforms baselines and stays robust when data is scarce or lower in quality. It gives no numbers.

## Why it matters

Safety rules are hard to write as cost functions, and demonstrations that are both safe and high-performing are costly to collect. This work suggests a cheaper route: mix two imperfect data sources. That could make safe robot learning easier to set up in real settings. The abstract does not say how the safety guarantee holds for the practical algorithm rather than the ideal solution.

## How it changes the race

If the results hold, teams could train safer robot policies from data they already have, without hand-built cost functions or perfect demonstrations. This is a gain for safe offline learning in robotics, not for general frontier SI capability.

## Safety and control

The paper bears directly on safety. It aims to keep agents out of unsafe regions while still using high-return behavior. The proof covers the ideal exact solution, and the abstract does not say that the trained agent carries the same guarantee.

### The abstract, by the authors

Ensuring high returns while providing strong safety guarantees in offline imitation learning (IL) is a fundamental challenge: existing approaches either rely on explicitly specified cost functions, which are difficult to design for complex real-world safety constraints, or require expert demonstrations that are both safe and high-performing, which are often too costly to collect. We introduce dual-expert imitation learning, which leverages two complementary data sources: (i) safety-compliant demonstrations that faithfully follow safety guidelines but may be suboptimal, and (ii) performance-oriented demonstrations that achieve high returns but may violate safety guidelines. To address this problem, we propose DexDICE (Dual-expert offline imitation learning via stationary DIstribution Correction Estimation). Instead of conventional distribution constraints, DexDICE combines DICE-style offline learning with safe support constraints, enabling agents to exploit high-return behaviors while remaining within safety-compliant regions. We prove that the ideal solution satisfying this constraint exactly has zero unsafe-region visitation with high probability. We evaluate DexDICE on a re-curated DSRL benchmark and validate it on a real-world mobile-robot navigation task trained directly from human teleoperation demonstrations, where it substantially outperforms baselines and remains robust to data scarcity and quality degradation.

Quoted from arXiv:2611.09433. The research and its words belong to its authors.

Desk score **77** out of 100

Why it was picked

- Safety (in the title)
- Agents
- Evals
- Security
- Accepted at NeurIPS
- Cross-listed in 2 fields
- Code or data released

Topics

- [Safety](https://thesimind.si/papers/?tag=safety)
- [Robotics](https://thesimind.si/papers/?tag=robotics)
- [Reinforcement learning](https://thesimind.si/papers/?tag=reinforcement-learning)
- [Training](https://thesimind.si/papers/?tag=training)
- [Alignment](https://thesimind.si/papers/?tag=alignment)
- [Benchmarks](https://thesimind.si/papers/?tag=benchmarks)

**arXiv ID:** 2611.09433v1
**Submitted:** November 6, 2026
**Fields:** cs.LG (Machine Learning), cs.AI (Artificial Intelligence)
**Level:** Advanced
**Comments:** NeurIPS 2026. 55 pages, 13 figures. Code: [https://github.com/tzs930/dexdice](https://github.com/tzs930/dexdice)

Fingerprint, SHA-256

`98912053547fed9b2bb5dd4f8aca3f39e48d809cdc83cc5b4881072f7766130e`

On Solana: [Proof of Mind seal, 437gF7...BR92](https://solscan.io/tx/437gF725DAG8gRKPBLFDmKiCEc8recBwc9f2nJbDSHirzNkcCnPsgSDCZzxXS15qHXCpomLuKrJuXUTtZbKhBR92)

### The exact text fingerprinted

{"arxiv":"2611.09433","version":1,"title":"DexDICE: Reconciling Safety and Performance via Dual-Expert Offline Imitation Learning","authors":["Seokin Seo","Sung Kuk Shyn","Minji Seo","Yisoo Lee","Jongmin Lee","Hongseok Yang","Kee-Eung Kim"],"day":"2026-11-09","rank":8,"score":77,"status":"written","summary":{"takeaway":"DexDICE learns from two kinds of demonstrations, safe but weaker ones and strong but risky ones, to build agents that score well while staying inside safe regions.","found":"The authors study offline imitation learning, where an agent learns only from recorded demonstrations. They propose \"dual-expert\" learning with two data sources: safety-compliant demonstrations that may be suboptimal, and performance-oriented demonstrations that may break safety guidelines. Their method, DexDICE, combines DICE-style offline learning (which corrects for differences in how often states are visited) with safe support constraints. These constraints keep the agent inside regions the safe data covers. They prove that the ideal solution, if the constraint is met exactly, has zero unsafe-region visitation with high probability. They test it on a re-curated DSRL benchmark and on a real mobile-robot navigation task trained from human teleoperation demonstrations. The abstract says it substantially outperforms baselines and stays robust when data is scarce or lower in quality. It gives no numbers.","matters":"Safety rules are hard to write as cost functions, and demonstrations that are both safe and high-performing are costly to collect. This work suggests a cheaper route: mix two imperfect data sources. That could make safe robot learning easier to set up in real settings. The abstract does not say how the safety guarantee holds for the practical algorithm rather than the ideal solution.","race":"If the results hold, teams could train safer robot policies from data they already have, without hand-built cost functions or perfect demonstrations. This is a gain for safe offline learning in robotics, not for general frontier SI capability.","safety":"The paper bears directly on safety. It aims to keep agents out of unsafe regions while still using high-return behavior. The proof covers the ideal exact solution, and the abstract does not say that the trained agent carries the same guarantee.","level":"advanced","tags":["safety","robotics","reinforcement-learning","training","alignment","benchmarks"]}}

[Check it on Verify a thought](https://thesimind.si/proof-of-mind/)

## The authors

Seokin Seo, Sung Kuk Shyn, Minji Seo, Yisoo Lee, Jongmin Lee, Hongseok Yang, Kee-Eung Kim

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.

Related on the desk

## More papers on these topics

[Nov 9, 2026 · Number 2Forcing a Letter Suppresses Refusal: Mechanistic Evidence from WMDPIn multiple-choice safety tests, forcing a model to pick a letter seems to crowd out its refusal, so these scores may measure knowledge more than safety.SafetyInterpretabilityEvals](https://thesimind.si/papers/2611.08639/) · [Nov 9, 2026 · Number 4HomeLoop: A Simulation Framework and Benchmark for Robot-IoT Coordination in Everyday Domestic Assistance and SafetyHomeLoop is a new simulated home benchmark where SI robots must work with smart devices and handle ambiguous alerts. The best of six models scored 61.9%, and models often over-reacted.RoboticsBenchmarksEvals](https://thesimind.si/papers/2611.08914/) · [Nov 9, 2026 · Number 6Alignment by Confident Debate: Preference Learning from Multi-Agent DebateAbCD turns debates among several SI models into training preferences without gold answer labels, and the authors report better accuracy than ensembling, debate alone, or SFT on the same traces.AlignmentReasoningMulti-agent systems](https://thesimind.si/papers/2611.09568/)

[Number 7](https://thesimind.si/papers/2611.09349/) · [All 12 for Nov 9, 2026](https://thesimind.si/papers/day/2026-11-09/) · [Number 9](https://thesimind.si/papers/2611.09082/)