In plain words
Some language models receive graph information twice: as a learned token and as words describing the same graph. That makes it hard to tell whether the token helps answer questions.
The authors tested whether these tokens differ across examples and whether changing them changes answers. In the default Llama-2 setup, tokens were nearly alike, and replacing them did not reliably change answer scores while the graph description remained. A two-stage training procedure made tokens more distinct, but this did not always improve scores.
The research
When a model receives graph information both as a learned token and as a written description, which channel actually supports its answers? Answer accuracy alone cannot distinguish them, so a strong score may leave the token’s role unclear.
The authors audit graph-token variation and use, then train a two-stage system that reduces near-identical tokens without guaranteeing better answers.
The authors ask whether a graph token helps answer questions when the same graph is also written into the prompt. To separate the channels, they replace tokens at inference while keeping the model fixed and the written graph present, then compare answer scores with the original-token condition.
For default G-Retriever with frozen Llama-2, the substitutions did not produce a resolved score change under the authors’ threshold. A graph-independent token trained through the same projector matched or exceeded its scores on all three tests.
This supports the conclusion that graph-specific token variation was not needed for those scores; the written graph remained available.
The detailed tests focus on G-Retriever on three graph question-answering tests; the audit of later methods covers selected runnable systems on WebQSP.
The proposed pipeline is a diagnostic proof of concept, and replacing a graph tests graph-question pairing rather than whether the model reasons over each graph fact.
Analysis
Deep
- Original paper authors
- Andrea Giuseppe Di Francesco, Rishabh Jain, Franco Maria Nardini, Nicola Tonellotto, Pietro Liò
- Submitted to arXiv
- License
- CC BY 4.0