Explain BenchFill with a conceptual diagram

Replace the numerical results chart with a schematic showing how filling difficulty gaps makes localized training gains easier to detect.

README.md

@@ -43,7 +43,7 @@ BenchFill generates new items to fill gaps in a benchmark's difficulty distribut
 
 **BenchFill helps detect training gains that existing benchmarks miss.** Filling gaps in benchmark difficulty makes localized improvements easier to observe. In the paper's SFT experiment, adding just one BenchFill question to each of three AIME benchmarks made training gains statistically detectable in all three cases; the original benchmarks and baseline augmentations detected none ([Table 12](https://naicheng-yu.github.io/Benchfill-Website/static/pdf/benchfill_paper.pdf#page=33)).
 
-![SFT accuracy gains: BenchFill detects significant gains in all three AIME experiments, compared with none for the original benchmarks or baseline augmentations.](benchfill-sft-results.png)
+![How BenchFill works: training gains in a difficulty gap may be missed by the original benchmark; adding questions in that gap makes the gains easier to detect.](benchfill-overview.png)
 
 ## Subsets