Do We Require Large Language Models To Be HHH Aligned For Reasoning, Safety, Value, and Pluralistic?
Abstract
AbstractAlignment in Large Language Models (LLMs) studies whether model responses satisfy the Helpful, Harmless, and Honest (HHH) objectives. Imbalances across these objectives may produce divergent behavior across Reasoning, Safety, Value, and Pluralistic dimensions. Existing benchmarks including CR-LLM, SAFETUNEBENCH, VALUEBENCH, and PERSONA study these dimensions separately and capture partial aspects of HHH behavior. To study HHH behavior across overlapping dimensions, we introduce RSVP-HHH, a framework for HHH evaluation across all four dimensions. We construct RSVP-X, an English-HHH dataset containing 9,228 prompt-response pairs spanning 106 domains through Step 1 Latent Behavioral Projection, Step 2 Behavioral Region Overlap, Step 3 Density-Aware Downsampling, Step 4 Behavioral Compatibility Evaluation, and Step 5 Latent Refinement. Experimental results on RSVP-X show that TrinityX (46.6%) outperforms AlignX (45.0%), SAGE (49.1%) outperforms VAL-BENCH (48.2%) and R-Zero (47.6%), DeepSeek (45.0%) outperforms Qwen (43.7%), and Gemini (49.6%) outperforms Claude (47.7%) in Avg across HHH objectives.