Research map / Large language models / Instruction and fine-tuning

Training-Free Task Vectors for LLM Behavioral Control

NeurIPS 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Training-Free Task Vectors (TFTVs), which map contrastive activation steering vectors to rank-one weight edits using SVD of the module weight and expected input statistics, enabling persistent, compositional behavioral control without fine-tuning.

Problem

Task vectors require a fine-tuned checkpoint to extract weight-space behavioral directions, making post-training control of LLM traits costly; activation steering gives such directions but only transiently at inference.

Approach

Maps a normalized steering vector to a rank-one weight update, using the left factor as the steering direction and the right factor as an SVD-weighted, sign-corrected vector aligned with the module's expected input, so the update needs only forward-pass statistics and is linear in steering vectors.

Key results

• On Llama-3.1 with trait addition, TFTV reaches 63.26 evil, 98.55 hallucination and 94.64 sycophancy trait scores, exceeding the best training-free baseline in five of six settings while keeping Llama MMLU within 0.15 points of base.
• For suppression, TFTV reduces trait scores by 7.74 to 77.28 points relative to inference steering and achieves the strongest suppression among training-free methods in most settings.
• In three-way composition on Llama-3.1, TFTV keeps MMLU at 68.36 and GSM8K at 79.15 while reducing sycophancy to 1.94 (base 92.06).

Surprising findings

• Composing edits sometimes improves GSM8K above the single-trait edits (e.g. S+H gives GSM8K 80.36 vs 78.54 base on Llama).
• Steer2Edit collapses under composed edits on Qwen 2.5, with coherence approaching zero for several settings.

Limitations

• Performance depends on which modules and layers are edited; automatic selection is left to future work.
• Trade-off between trait control and utility persists under stronger or composed edits.
• Composition is only approximately linear when expected inputs are estimated separately per trait.

Experimental setup

Datasets: Persona Vectors evaluation questions and trait-eliciting system prompts Moral Stories TruthfulQA MMLU GSM8K

Benchmarks: Persona Vectors evaluation (evil, hallucination, sycophancy) MMLU GSM8K Moral Stories TruthfulQA sycophancy datasets (NLP, philosophy, politics)

Metrics: trait score coherence score MMLU accuracy GSM8K accuracy MC1 MC2

Baselines: Persona Vectors steering Steer2Edit Task Vectors Contrastive Weight Steering

Builds on: task vectors Persona Vectors Steer2Edit Contrastive Weight Steering

Closest papers by meaning

• Efficient Model Editing with Task-Localized Sparse Fine-tuning (ICLR 2025 · similarity 0.89)
• Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.88)
• When is Task Vector Provably Effective for Model Editing? A Generalization Analysis of Nonlinear Transformers (ICLR 2025 · similarity 0.87)
• Steer2Edit: From Activation Steering to Component-Level Editing (NeurIPS 2026 · similarity 0.87)
• Improving Instruction-Following in Language Models through Activation Steering (ICLR 2025 · similarity 0.86)
• QueST: Self-Supervised Skill Abstractions for Learning Continuous Control (NeurIPS 2024 · similarity 0.86)