On the Token Collapse Phenomenon in Textual Graph Soft Prompting
3 judges. Scored by all three judges. Only the day’s top stories get the claim check. How far stories are checked
The paper analyzes G-Retriever, a method that feeds graph-neural-network soft tokens alongside a textual graph description into an LLM, and finds a failure it calls token collapse: the graph tokens become nearly identical across inputs.
Score
0.63total under rubric v2.0
- Significance0.57
- Reader interest0.65
- Audience breadth0.38
- Practical relevance0.61
- Substantiated0.87
- Technical depth0.81
- New information0.93
- Counts against: Promotional0.03
- Counts against: Overclaim0.17
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| Significance | 0.60 | 0.40 | 0.72 | 0.57 | 0.32 | +0.25 | +0.144 |
| Reader interest | 0.51 | 0.55 | 0.88 | 0.65 | 0.37 | +0.20 | +0.129 |
| Audience breadth | 0.43 | 0.30 | 0.42 | 0.38 | 0.13 | +0.10 | +0.038 |
| Practical relevance | 0.62 | 0.50 | 0.70 | 0.61 | 0.20 | +0.15 | +0.091 |
| Substantiated | 0.98 | 0.80 | 0.82 | 0.87 | 0.18 | +0.10 | +0.087 |
| Technical depth | 0.76 | 0.80 | 0.86 | 0.81 | 0.10 | +0.10 | +0.081 |
| New information | 0.95 | 0.85 | 0.99 | 0.93 | 0.14 | +0.10 | +0.093 |
| Promotional | 0.05 | 0.03 | 0.01 | 0.03 | 0.04 | −0.15 | −0.004 |
| Overclaim | 0.17 | 0.20 | 0.14 | 0.17 | 0.06 | −0.15 | −0.026 |
| Sum of contributions | +0.632 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.632 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN —
Recorded for context only, checked 24 Oct 2026, 08:54 UTC. It played no part in the ranking.