Improving Quantized Model Performance in Qualitative Analysis with Multi-Pass Prompt Verification

Technical Summary: Improving Quantized Model Performance in Qualitative Analysis with Multi-Pass Prompt Verification

Problem Statement
Large Language Models (LLMs) are increasingly utilized for qualitative analysis due to their ability to summarize texts and identify themes efficiently. However, reliance on full-precision models or commercial APIs introduces significant computational costs and privacy risks. While quantization (reducing bit precision of weights and activations) offers a solution for resource-constrained local deployments, it often degrades model performance. Specifically, low-bit quantization (e.g., 2-bit to 4-bit) increases hallucinations, causes semantic drift, and reduces accuracy, particularly when processing non-expert language characterized by unclear or inconsistent terminology. Existing research has addressed hallucination reduction and prompt design in isolation but lacks a systematic evaluation of how varying quantization levels and types affect LLM performance in qualitative tasks, nor does it offer a framework to stabilize low-bit models for such analyses.

Methodology
The study evaluates the LLaMA-3.1 (8B) model across a dataset of 82 semi-structured interview transcripts (ranging from 8,000 to 13,000 words) involving both expert and non-expert participants. The methodology follows a five-phase workflow:

1. Ground Truth Construction: Two researchers independently coded transcripts using NVivo. These outputs were combined with a corrected BF16 LLaMA-3.1 output (manually verified) to create a Gold-Standard Ground Truth (GSGT). This process ensured the baseline was rigorously validated and free from circularity or model bias.
2. Quantized Evaluation: The study tested multiple quantization configurations:
   • 8-bit: SmoothQuant, INT8, GGUF Q8_0.
   • 4-bit: Q4_K_M, NF4, GPTQ-4bit.
   • 3-bit: HQQ, AWQ.
   • 2-bit: SPQR, GGUF Q2_K.
     Models were evaluated both "Before" (single-pass generation) and "After" (applying the proposed framework).
3. Proposed Framework: A Quantization-Aware Multi-Pass Prompt Verification method was introduced. This framework operates in controlled steps:
   • Analysis Pass: The model extracts themes, sub-themes, and frequencies using structured system prompts and JSON constraints to reduce cognitive load.
   • Verification Pass: A second prompt instructs the model to verify the initial output against the original transcript. It removes unsupported themes, validates quotes, and corrects frequency counts.
   • Iteration: The process repeats until convergence (typically within two passes), ensuring only transcript-grounded content is retained.
4. Metrics: Performance was measured using eight metrics: Theme Extraction F1-Score, Semantic Drift Score (SDS), Hallucination Rate (HR), Theme Consistency Score (TCS), Frequency Correlation, Keyword Omission Rate, Keyword Hallucination Rate, and Adjusted Rand Index (ARI).

Key Contributions

1. Systematic Evaluation: The paper provides the first comprehensive evaluation of multiple quantization levels (2-bit to 8-bit) and specific quantization types (e.g., AWQ, SPQR, GPTQ) specifically for qualitative transcript analysis, focusing on hallucination and thematic alignment rather than generic reasoning benchmarks.
2. Verification Framework: It introduces a quantization-aware multi-pass verification workflow designed specifically to stabilize low-bit local LLM deployments. This framework isolates the effects of quantization from knowledge augmentation by enforcing strict internal transcript validation.
3. Empirical Validation: The study demonstrates that structured verification can compensate for precision loss in resource-constrained settings, enabling practical qualitative analysis without increasing model size or requiring full-precision hardware.

Results

• Baseline Performance: Without verification, low-bit models (3-bit and 2-bit) exhibited significant performance degradation, with high hallucination rates and low thematic consistency, especially on non-expert inputs. 8-bit models remained closest to the GSGT but still showed minor drift.
• Impact of Verification: The multi-pass framework significantly improved performance across all quantization levels:
  • 4-bit Models: Showed the most dramatic relative gains (40–80% improvement), becoming competitive with 8-bit baselines while requiring far less memory.
  • 3-bit and 2-bit Models: While absolute accuracy remained lower than higher-bit models, the framework substantially reduced hallucinations and stabilized outputs, making them viable for structured analysis where they were previously unusable.
  • Expert vs. Non-Expert: Non-expert inputs (unclear language) suffered more from quantization but benefited most from the verification process, which effectively removed unsupported themes and corrected frequency counts.
• Component Analysis: Ablation studies indicated that theme-level verification was the primary driver of hallucination reduction, while quote validation improved semantic grounding, and frequency verification enhanced quantitative reliability.

Significance
The paper claims that verification efficiency is a critical factor in enabling reliable low-resource LLM deployments, often outweighing the benefits of increased model precision alone. By demonstrating that a structured, multi-pass verification workflow can mitigate the specific errors introduced by aggressive quantization (such as semantic drift and hallucination in non-expert text), the study enables faster, more accurate, and privacy-preserving qualitative research. The findings suggest that low-bit quantized models, when paired with this verification framework, can achieve performance levels suitable for practical use, democratizing access to advanced qualitative analysis tools without the computational overhead of full-precision models.