REDDIT · REDDIT// 3d agoOPENSOURCE RELEASE

Sarvam reasoning models hit abliteration

Researcher Alosh Denny removes refusal mechanisms from India’s first multilingual MoE reasoning models, Sarvam-30B and 105B. The "abliteration" process reveals a unique dissociation between internal reasoning and final answer projection, showing that reasoning models can "think" toward compliance while still refusing in the output.

// ANALYSIS

Reasoning models possess a dual-circuit refusal architecture that complicates safety alignment beyond standard LLM techniques.

–Identified two distinct refusal circuits: one in core reasoning layers and one at the final lm_head projection.
–Refusal mechanisms are pre-linguistic; English-computed directions successfully uncensored Malayalam, Hindi, and Kannada outputs.
–Uncensored 105B variant remains highly competitive, maintaining 98.6 on Math500 and 71.7 on Live Code Bench.
–Surgical Refusal Ablation (SRA) proves that safety layers in reasoning-heavy MoEs are structural rather than language-dependent.

// TAGS

llmreasoningopen-weightsmultilingualsarvam-aisafetyabliteration

DISCOVERED

3d ago

2026-04-09

PUBLISHED

3d ago

2026-04-08

RELEVANCE

8/ 10

AUTHOR

Available-Deer1723