tiny-aya-subspace-steering-vectors
steering Tiny-Aya behavior without fine-tuning weights
3 min readupdated
A collection of steering vectors (directions in an AI model's internal activation space used to adjust outputs on the fly during generation) for Cohere's multilingual Tiny-Aya models (Global, Fire, and Water, 3.35B parameters). The authors tested whether multidimensional PCA subspaces outperform single vectors and found that, at equal intervention strength, a single well-chosen vector is often more effective and less likely to break coherence. As a result, they significantly reduced false refusals, sycophancy, and stereotypes without modifying the base model's weights. Useful for researchers and developers who need to fine-tune the behavior of compact LLMs (language models) for specific tasks.
Key features
- Sharp drop in false refusals: overall accuracy on
XSTestandOR-Benchjumped from 55.3% to 78.2% (+23.0 points), while the model still blocks 98.7% of genuinely unsafe requests. - Curbing sycophancy: principled responses on political topics in English rose from 30.5% to 72.0% (+41.5 points), and vectors trained only on English and Spanish transferred zero-shot to Japanese and Bengali.
- Prompt injection and bias resilience: system prompt hierarchy adherence increased from 30.7% to 43.9% (protection against user overrides in
SystemCheckwent from 0% to 40%), while stereotype bias onBBQwas cut in half (from 16.3 to 8.2). - Two intervention modes: the authors showed that projection (erasing unwanted signals) works best for suppressing false positives, while addition (injecting an attention vector) works best for catching contradictions in text.
What you can do
- Cut unnecessary censorship in assistants: attach projection vectors to
Tiny-Ayaso the model stops blocking benign queries containing trigger words like "kill a process in Linux". - Boost attention to detail in documents: use additive vectors in RAG pipelines (retrieving answers from knowledge bases) to help the model spot contradictions between text passages.
- Tune multilingual bots: apply cross-lingual transfer of steering vectors to stop the model from agreeing with false user claims across different languages.
Example
The model card outlines an inference-time setup for the Tiny-Aya family: hooks with 1D vectors or subspace bases are attached to the hidden states of selected layers. To suppress false refusals, a projection operation subtracts the component along the refusal axis; to detect text discrepancies (FaithEval) and protect the system prompt, an additive activation shift is applied while monitoring the cosine similarity of the next-token distribution (kept at 0.90 or above to preserve text coherence).
Read more
DetailsBase models: Cohere Tiny-Aya (Global, Fire, Water, 3.35B parameters).
Benchmarks: XSTest + OR-Bench — 78.2% vs. 55.3% baseline; FaithEval + TruthfulQA — contradiction detection rose from 54.6% to 72.7%; Multi-IF (8 languages) — from 23.8% to 33.8% with 5D PCA.
Highlights: datasets were deduplicated against leakage using MinHash (CuratorKIT), and shift strength was precisely calibrated per layer for fair method comparison.










Comments