AI Science Tracker
EN

South China University of Technology

2 discoveries from 2021 to 2026

Discoveries

A Bounded Evaluation Protocol for Representational Systematicity

Source check: Source checked. Not independent scientific confirmation.

Finding / tested result: On a controlled semantic-parsing testbed, the protocol classified role–filler binding as causal-use and the declared geometry explanation as unsupported within its searched space, with calibration and seed checks.

Technology: Bounded-RS protocol with matched counterfactual tests and causal interventions on fine-tuned T5 models

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Bounded-RS protocol with matched counterfactual tests and causal interventions on fine-tuned T5 models
Limitation / still unknown
The verdict is bounded to the specified explanations, models, testbed and search space.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/html/2611.01755v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 22 Nov 2026. Review scope: Primary result-bearing material for the stated method, result and evaluation, with bibliographic version information..

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Trained recurrent model preserves cue geometry across a memory delay

Source check: Source checked. Not independent scientific confirmation.

Finding: A trained256-unit recurrent model of context-dependent decisions retained cue-specific population geometry across a delay despite changing single-unit activity. Numerical subspace analysis linked low-activity delay states with later context-specific integration. This is a conditional computational interpretation, not a newly observed biological memory mechanism.

Technology: Rate-based recurrent neural network

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Training the recurrent weights for a delayed contextual task creates the dynamics being analyzed; epoch-specific subspace analysis tests how learned cue representations persist and guide integration.
Test and result
Original model/training Methods and cue-delay/manifold Results were read. The model learned color-versus-motion context rules, and numerical trajectories and fixed/slow points were analyzed within its fitted dynamics.
Limitation / still unknown
Known mixed-selectivity, dynamic-memory and low-activity theories are substantial precedents, limiting novelty. The analysis concerns a trained artificial model, without new neural recordings or independent biological validation. Dale-type excitatory/inhibitory constraints and biological plasticity were not implemented; numerical line-attractor analysis is not a formal proof or a universal account of working memory.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC8367843/ · https://doi.org/10.1016/j.isci.2021.102919

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 21 Nov 2026. Review scope: Actual relevant primary computational/clinical Methods, model/result context and limitations read; no operational biological protocols.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 1/5 · Validation 3/5 · Novelty 1/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open the interactive profile for the timeline, or browse all organizations and model families.