DHSA Achieves DPO-Level Alignment at Inference Time Using Only 256-16384 Control Parameters

1 reports archived1 independent sourcesupdated 10/12/2026, 12:05:00

Synthesis & Latest Updates

1 Sources Cross-Validated

Researchers propose Direct Hidden State Alignment (DHSA), which uses Residual Competition Maps (RCM) to quantify internal computational forces supporting or opposing preferences, enabling inference-time intervention without modifying base weights. With only 256-16384 control parameters, DHSA achieves alignment performance comparable to DPO while preserving base model diversity, highlighting the distinction between aligned outputs and consistent internal mechanisms.

LATEST/Researchers propose Direct Hidden State Alignment (DHSA), which uses Residual Competition Maps (RCM) to quantify internal computational forces supporting or opposing preferences, enabling inference-time intervention without modifying base weights. With only 256-16384 control parameters, DHSA achieves alignment performance comparable to DPO while preserving base model diversity, highlighting the distinction between aligned outputs and consistent internal mechanisms.

HEAT TRENDHourly heat curve

2 fully observed hours

Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first

量子位 (微信公众号)T2·Oct 12, 12:0578 pts
DHSA Achieves DPO-Level Alignment at Inference Time Using Only 256-16384 Control Parameters
- Introduces Residual Competition Maps (RCM) to map external preferences back to native inference computations, quantifying whether components support or oppose target behaviors.
- Reveals that even when model outputs align with preferences, internal opposing computational forces may persist; methods like DPO do not fully eliminate this internal competition.
- DHSA achieves DPO-level alignment via inference-time hidden state intervention using minimal control parameters, without modifying base model weights.
View Evidence Original Source