On the Token Collapse Phenomenon in Textual Graph Soft Prompting

On the Token Collapse Phenomenon in Textual Graph Soft Prompting. By Andrea Giuseppe Di Francesco and five others, from Sapienza University of Rome, the Institute of Information Science and Technologies of the Italian National Research Council, and two other institutions. Published on the twenty-second of October. Its code repository is worked out: about forty-five source files, with a readme.

One recent way to bring graph structure into a language model's prompt is graph soft prompting. A graph neural network, a GNN, learns graph-specific tokens meant to capture the relations among the entities a textual prompt mentions. G-Retriever pairs those graph soft prompts with a verbalized description of the same graph, to strengthen the textual signal.

Looking closely at G-Retriever, the authors find a pathology they call token collapse. The graph tokens, which are supposed to differ from graph to graph, become almost indistinguishable from one input to the next. Their conclusion is that the collapsed token is not carrying graph content as intended. It works mainly as a generic soft prompt.

The collapse is not confined to one setup. It appears across three benchmarks, three graph encoders, two families of language model, and with the model both frozen and LoRA-adapted. The authors also give evidence that methods which came after G-Retriever inherit the same failure, even when they pool several tokens.

Two kinds of test show the token is not doing graph work. In the first, under the standard setup, the token is swapped out: for the dataset mean, for a token taken from another example, or for one built from a corrupted graph. None of these swaps produces a resolved effect on answer accuracy. If the token were carrying the graph, putting in the wrong graph, or a broken one, ought to matter. It does not.

In the second test, the graph is taken away entirely. A learned constant prompt, one that never accesses the graph at all, matches or beats G-Retriever on all three benchmarks. Whatever the graph token was contributing to accuracy, a fixed prompt with no graph behind it contributes as much.

The paper goes on to examine conditions linked to the collapse, and proposes a two-stage procedure that yields tokens which do not collapse, while keeping performance competitive in most cases. Most, not all.

A second limit is drawn just as clearly. Tokens that vary are not proof that the model uses graph information. Using interventions and decoding probes, the authors show that, depending on the configuration, a token can reflect the question, an answer prior, or content derived from the graph, and in that last case the language model may or may not make use of it. Restoring variability does not, by itself, settle whether the graph is being used.

Before this, a graph soft prompt that helped accuracy was taken to be delivering the graph. After it, accuracy alone is not enough evidence: a constant prompt can score as well, and whether a graph token carries the graph has to be checked directly. The authors offer their tests as a practical way to evaluate that alignment between graph and token, and they release code to reproduce the experiments.