Shortcuts and Boundaries: Verifiable RL Misalignment and Contextual Safety Learning
cs.AI, cs.CR
Submitted: 2026-11-09
Updated: 2026-11-11
Terminology
Sources
- Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms
- ST-WebAgentBench: A Benchmark for Evaluating Safety and Trustworthiness in Web Agents
- Identifying the Risks of LM Agents with an LM-Emulated Sandbox
- $\tau$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains
- LPS-Bench: Benchmarking Safety Awareness of Computer-Use Agents in Long-Horizon Planning under Benign and Adversarial Scenarios
- SEAD: A State-Based Perspective on Attack and Defense in Tool-Using Agents
- BLINDSPOT: A Benchmark for Safety and Refusal Calibration in Long-Horizon Tool-Using Agents
- The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes
- BAITBENCH: Measuring Agent Reward Hacking with Optional Shortcuts Planted in ML Tasks
- The Trojan Knowledge: Bypassing Commercial LLM Guardrails via Harmless Prompt Weaving and Adaptive Tree Search
- One Turn Too Late: Learning When to Intervene Against Multi-Turn Malicious Intent
- Deliberative Alignment: Reasoning Enables Safer Language Models
- Defining and Characterizing Reward Hacking
- Scaling Laws for Reward Model Overoptimization
- Hack-Verifiable Environments: Towards Evaluating Reward Hacking at Scale
- BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure
- Constitutional AI: Harmlessness from AI Feedback
Related papers
- Multi-Objective Structured Pruning of LLMs for Latency and Model Size Optimization
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Embeddings for Preferences, Not Semantics
- Can We Trust Item Response Theory for AI Evaluation?
- Diagnosing Task Insensitivity in Language Agents