MultiSoc-4D: A Benchmark for Diagnosing Instruction-Induced Label Collapse in Closed-Set LLM Annotation of Bengali Social Media

Authors: Souvik Pramanik, S. M. Riaz Rahman Antu, Shak Mohammad Abyad et al.

arXiv · PDF

Summary

The authors built MultiSoc-4D, a 58,000-comment Bengali social-media dataset annotated for category, sentiment, hate speech, and sarcasm, and used it to diagnose a systematic LLM annotation failure they call "instruction-induced label collapse." When asked to label closed-set categories (e.g., hateful/not hateful), ChatGPT, Gemini, Claude, and Grok all showed strong bias toward safe fallback labels ("Other," "Neutral," "No"), missing 79% of hateful content and 75% of sarcasm compared to human-calibrated labels. Even though the models agreed with each other at high rates, Fleiss' kappa was near zero for sarcasm, revealing what the authors call a "label agreement illusion"—models converge on the wrong answer together.

Main takeaways:

• LLMs systematically prefer neutral or "Other" labels in closed-set annotation tasks, especially for minority categories like hate speech and sarcasm.
• High inter-LLM agreement doesn't guarantee quality: the models can all collapse to the same wrong default, producing near-zero kappa despite surface consensus.
• The effect persists across 40+ LLMs of different architectures, suggesting it's a widespread training-pipeline issue rather than model-specific.
• The dataset is released as a diagnostic benchmark for annotation bias in low-resource (Bengali) NLP.
• The bias propagates downstream: models trained on LLM-annotated data inherit the label collapse.

Relevance

Tangentially relevant to my work on conditional behavior and emergent misalignment: this shows that LLMs have strong default-label tendencies that override instruction, which could be a related phenomenon to persona leakage or marker uptake—both involve competition between instruction and learned priors.

Threat model

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses bias, benchmark.

Abstract

arXiv:2605.06940v1 Announce Type: new Abstract: Annotation automation via Large Language Models (LLMs) is the core approach for scaling NLP datasets; however, LLM behavior with respect to closed-set instructions in low-resource languages has not been well studied. We present MultiSoc-4D, a Bengali social media dataset benchmark, which contains 58K+ social media comments from six sources annotated along four dimensions: category, sentiment, hate speech, and sarcasm. By employing a structured pipeline where ChatGPT, Gemini, Claude, and Grok individually annotate separate partitions, while sharing a common validation set of 20%, we diagnose LLM behavior systematically. We discover a prevalent phenomenon called "instruction-induced label collapse", wherein LLMs show a systematic preference towards fallback labels (Other, Neutral, No), leading to high agreement rates but under-detection of minority categories. For example, we find that LLMs failed to detect 79% and 75% of instances with hateful and sarcastic content compared to a human-calibrated reference. Furthermore, we prove that it represents a "label agreement illusion", statistically validated via almost null Fleiss' Kappa ($\kappa \approx -0.001$) on sarcasm detection. Across 40+ LLMs, we benchmark this annotation bias propagation within the training pipeline, regardless of architectural differences. We release MultiSoc-4D as a diagnostic benchmark for annotation biases in Bengali NLP.