Original title: From Confusion to Clarity: Confusion-Aware Retrieval and Knowledge Injection for Text Classification
Manish Gupta, Chaitanya Giri, Jayasimha Talur · September 2026
Large language models often struggle to sort text into categories that mean almost the same thing. This happens because the exact differences are specific to a subject and were never learned during early training. Usually, systems search for the top most relevant labels and ask the model to pick one. But when two very similar labels both show up, the model cannot tell them apart. Researchers built a new system to fix this without extra training. First, it spots which labels trick the model. Next, it adds those easily confused labels to the candidate list. Then, it writes specific rules showing how to tell them apart. This method improves accuracy, and the rules even help smaller, cheaper models make better choices.
Imagine sorting a pile of laundry into drawers for similar colors like navy blue and black, but you often mix them up because they look alike in certain lights. Normally, you might just pick a few drawers that seem right, but if both the navy and black piles are there, you might grab the wrong sock. In this paper, the system first notices which color pairs you confuse, then adds those to your shortlist and gives you a specific tip, like 'navy fades in sunlight while black stays dark,' to help you choose correctly when you're actually sorting.
When AI sorts text into categories with very similar meanings, it makes errors, affecting fields like law or product catalogs. This framework boosts accuracy without retraining models, making AI cheaper and more reliable for users working with fine-grained labels.
The abstract doesn't state explicit limitations, but the approach is tested on three benchmarks, so it may not work as well for all types of text or label structures. It also depends on correctly identifying confusable pairs, which could be limited by the data available.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
From Confusion to Clarity: Confusion-Aware Retrieval and Knowledge Injection for Text Classification — Manish Gupta, Chaitanya Giri, Jayasimha Talur
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.