GLiClass: lightweight classification with dynamic candidates

简体中文

Paper information

Field	Details
Paper	arXiv 2508.07662
Organization	First-author affiliation: Knowledgator
First public release	2025-08-11 (arXiv v1)
Authors' code	Knowledgator/GLiClass
Adapter / method	system-one:bilinear-ce, system-one:rival-ce
Local reproduction code	src/auto_research/system_one/

Original paper

Motivation and method

GLiClass supplies text and runtime candidate labels to a single classification interface. This avoids unstable parsing of generative LLM outputs and separate cross-encoder inference on each text–label pair. The repository retains the defining mechanisms: candidate content affects scores, and one normalization produces a distribution over all candidates.

flowchart LR
  S[State text] --> E[Text encoding]
  C[Dynamic candidates and descriptions] --> L[Candidate encoding]
  E --> X[Bilinear or rival-centered interaction]
  L --> X
  X --> P[Softmax probabilities and selection]

The local NumPy mechanism baseline does not load GLiClass weights or reproduce the full Transformer, PPO multi-label training, or paper results. It is used in the shared System One Banking77 protocol and evolve operator.

Key figure from the paper

Figure 1: Core architecture, modules, and their connections. From the original paper; copyright remains with the authors.

Core computation

The local scorer computes softmax(h_state W h_candidate). The rival variant first centers candidate representations within the same question, then scores them.

Paper results

The paper reports zero- and few-shot results across classification tasks. Local results use the shared Banking77 protocol, with no direct comparison against paper numbers from different model scales or hardware.

Reproduction boundaries

Only dynamic candidate interaction and single normalization are reproduced, not the full GLiClass weights, training corpus, or PPO pipeline.