MASteer: Adaptive LLM Trustworthiness Repair

• MASteer is a fully modular framework that uses adaptive representation engineering to diagnose and repair LLM trustworthiness issues.
• It automatically generates high-quality steer samples with AutoTester and employs AutoRepairer for context-specific, inference-time interventions.
• Experimental results show significant accuracy improvements on benchmarks while preserving general-purpose performance and efficiency.

MASteer is an end-to-end, training-free framework for LLM trustworthiness repair based on adaptive representation engineering. MASteer addresses limitations of prior LLM repair paradigms—such as supervised fine-tuning, RLHF, and prompt engineering—by offering automated steer sample generation and context-driven activation steering at inference. It comprises two principal subsystems, AutoTester and AutoRepairer, which together constitute a fully modular pipeline for identifying, characterizing, and remediating undesirable LLM behaviors across a variety of trustworthiness dimensions, including truthfulness, fairness, and safety (Li et al., 9 Aug 2025).

1. Motivation and Objectives

LLMs manifest persistent failures in trustworthiness—hallucinations, bias, and safety hazards—posing challenges to deployment in high-stakes environments. Existing remedies such as supervised fine-tuning and RLHF are resource-intensive and slow to adapt to new issues, while prompt engineering is brittle and lacks coverage. Contemporary representation engineering approaches provide lightweight interventions by injecting concept vectors at inference but depend on hand-engineered samples and static strategies. MASteer is designed to automate sample generation and strategy selection, creating a scalable and adaptive LLM repair pipeline that delivers robust improvements in task-specific trustworthiness while maintaining general-purpose capabilities.

2. System Architecture and Workflow

MASteer consists of two modular agents:

2.1 AutoTester: Multi-Agent Steer Sample Generation

AutoTester automates the construction of high-quality, contrastively labeled QA triplets ⟨q, a⁺, a⁻⟩ suitable for steering vector discovery. The process involves four interacting roles:

• Analyst: Decomposes the target issue $I$ into orthogonal categories $C=\{c_1, ..., c_K\}$ and defines granular test scopes $T_{c_i}$.
• Retriever: For each $(c_i, T_{c_i})$, retrieves real-world references $R_{c_i}$.
• Writer: Drafts candidate QA samples per reference and iteratively refines them.
• Reviewer: Evaluates Relevance, Steerability, and Learnability, enforcing rigorous filter criteria for inclusion in the final steer dataset $S$.

The process is formalized as an agent collaboration loop (see Algorithm 1 in (Li et al., 9 Aug 2025)); output is a large, orthogonally structured set $S$ tailored to developer-specified trustworthiness defects.

2.2 AutoRepairer: Adaptive Strategy Selection and Steering

AutoRepairer operationalizes representation engineering through adaptive selection of steer vectors, injection layers, anchor points, and strengths:

• Scholar: Maintains a library $\mathcal{A}$ of steering algorithms (e.g., Mean-Difference, PCA, Logistic Regression, K-Means).
• Proposer: For each $\mathcal{A}$ and each LLM layer $l$, computes steering vectors $C=\{c_1, ..., c_K\}$0 from positive/negative activations ($C=\{c_1, ..., c_K\}$1, $C=\{c_1, ..., c_K\}$2), and evaluates per-layer suitability via the weak sample ratio $C=\{c_1, ..., c_K\}$3.

Selection logic identifies the optimal intervention layer $C=\{c_1, ..., c_K\}$4 and partitions the sample set among candidate algorithms for per-context steering. Each strategy profile consists of an anchor vector $C=\{c_1, ..., c_K\}$5, a steer vector $C=\{c_1, ..., c_K\}$6, and an intervention strength $C=\{c_1, ..., c_K\}$7. At inference, per-input logic maximizes cosine similarity with anchor vectors to select the most appropriate steering strategy, which is then injected to modulate the LLM’s response in context.

3. Mathematical Formalism for Adaptive Representation Engineering

Steering in MASteer is governed by a modular, algorithm-agnostic mathematical foundation:

• Activation Injection: For layer $C=\{c_1, ..., c_K\}$8, let $C=\{c_1, ..., c_K\}$9 be the input post-residual, $T_{c_i}$0 the post-attention activation, and $T_{c_i}$1 the usual post-FFN output. Steering injects an affine modification:

$T_{c_i}$2

where $T_{c_i}$3 is the steer vector, and $T_{c_i}$4 controls intervention strength.

• Steer Vector Computation: For activation sets $T_{c_i}$5, canonical options include Mean-Difference (MD), PCA/RepE, logistic regression (“ITI”), and K-Means centroids:

$T_{c_i}$6

Additional methods build steer vectors as learned or geometric separators between contrast sets.

• Layer and Algorithm Selection: At each layer, difference activations $T_{c_i}$7 are scored; the “weak sample ratio” $T_{c_i}$8 quantifies discriminability, guiding choice of the optimal $T_{c_i}$9.
• Anchor and Strength Assignment: For each algorithm, the mean negative activation $(c_i, T_{c_i})$0 and default intervention strength $(c_i, T_{c_i})$1 are computed from sample assignments.
• Per-input Strategy Assignment: At inference, for input $(c_i, T_{c_i})$2, extract $(c_i, T_{c_i})$3 and select $(c_i, T_{c_i})$4, then inject $(c_i, T_{c_i})$5 with strength $(c_i, T_{c_i})$6.

4. Algorithmic Workflow

The MASteer pipeline is executed as follows:

1. Steer Sample Generation: The AutoTester loop (see Algorithm 1 in (Li et al., 9 Aug 2025)) produces a large batch of high-contrast, high-coverage steering samples $(c_i, T_{c_i})$7 matching issue $(c_i, T_{c_i})$8’s decomposition.
2. Strategy Synthesis:
   • For each $(c_i, T_{c_i})$9 and $R_{c_i}$0, extract $R_{c_i}$1, $R_{c_i}$2, compute $R_{c_i}$3.
   • Identify $R_{c_i}$4 (best separation in representation space).
   • Assign each sample to the algorithm maximizing cosine alignment.
   • Compute anchors $R_{c_i}$5 and strengths $R_{c_i}$6.
3. Inference-time Steering: For input $R_{c_i}$7, extract its relevant activation, select strategy via anchor proximity, and inject steer vector with calibrated strength.

This enables MASteer to operate fully inference-side, requiring no retraining or gradient updates on the base LLM.

5. Experimental Evaluation and Core Results

Extensive evaluation is performed on LLaMA-3.1-8B-Chat and Qwen-3-8B-Chat using standard and custom trustworthiness tasks:

• Benchmarks: TruthfulQA (truthfulness), BBQ (fairness), SafeEdit (safety), MMLU (general knowledge), AlpacaEval (alignment).
• Automatic Metrics: Multiple-choice accuracy (ACC) across tasks.
• MASteer Configuration: 1,000 steer samples per issue (10 categories × 10 scopes × 10 references), algorithm library $R_{c_i}$8 {MD, PCA, LR, K-Means}.

Results indicate that MASteer raises overall trustworthiness ACC on LLaMA-3.1-8B-Chat from 50.84% to 58.65% (+15.36%) and on Qwen-3-8B-Chat from 65.67% to 68.42% (+4.21%), also improving or maintaining performance on MMLU and AlpacaEval [(Li et al., 9 Aug 2025), Table 1]. In custom cases (e.g., enforcing formal-positive style), MASteer achieves up to 93.2% test ACC (versus 62% at baseline). The framework’s per-issue adaptivity and lack of training reliance are demonstrated by successful interventions on a wide range of harmful or unwanted outputs.

Model	TrustfulQA	BBQ	SafeEdit	MMLU	AlpacaEval
LLaMA (Baseline)	48.87	59.82	43.85	57.22	54.06
LLaMA+MASteer	56.30	66.54	53.11	61.21	57.13
Qwen (Baseline)	65.12	71.82	60.07	68.75	54.49
Qwen+MASteer	69.47	74.18	61.63	70.68	56.11

MASteer’s context-adaptive steering enables improvement on issue-specific metrics without general degradation.

6. Robustness, Generalization, and Deployment Characteristics

MASteer demonstrates stable improvements across diverse domains, supported by several robustness and generalization findings:

• Adaptive Strength and Strategy: Dynamic selection of both direction and magnitude ($R_{c_i}$9) outperforms fixed approaches; use of middle-layer intervention (e.g., LLaMA layer 13) captures abstract properties with limited off-target effects.
• General Capability Preservation: Intervention logic avoids unnecessary or harmful modifications, typically maintaining or slightly improving general metrics (MMLU, AlpacaEval).
• Sample Diversity & Customizability: The agent-based AutoTester enables high coverage for both standard and novel issue types, requiring only developer specification of the defect.
• Resource Efficiency: The approach is training-free, requires only inference-time computation (single GPU), and allows rapid, scalable deployment in production APIs and services.
• Extensibility: Scholar’s library is architected for routine extension to new vector computation methods, supporting continued advancement in steering techniques.

MASteer thus constitutes a modular and practically viable solution for automated, adaptive, and lightweight trustworthiness repair in modern LLMs (Li et al., 9 Aug 2025).