improve error handling in CFG

latex/mwpls2024/presentation_abstract.txt

@@ -0,0 +1,3 @@
+Title: Let's wrap this up! Incremental structured decoding with resource constraints
+
+Language models are known to struggle with few-shot constraint satisfaction tasks, such as responding within a fixed number of tokens or without using certain words. These scenarios typically arise in programming-related tasks such as code completion and program repair. We propose a novel incremental approach to constrained sequence generation based on a theory of Brzozowski (1964). Our approach constructs a valid sequence from left-to-right using next-token probabilities from the language model to steer the decoder towards natural solutions. We demonstrate the effectiveness of this technique on the problem of syntax repair, where it enjoys a significant advantage over state-of-the-art neural repair models in terms of precision on blind recovery of human repairs, while ensuring that every valid sequence is generable and every generable sequence is valid.

src/jvmMain/kotlin/ai/hypergraph/kaliningraph/parsing/ParallelSyntaxRepair.kt

@@ -121,16 +121,6 @@ fun bijectiveRepair(
   return if (parallelize) ::genSeq.parallelize().distinct().asSequence() else genSeq()
 }
 
-// This experiment essentially tries every possible combination of fillers in parallel
-fun List<Σᐩ>.parallelSolve(fillers: Set<Σᐩ>) =
-  MDSamplerWithoutReplacement(fillers, count { it == HOLE_MARKER })
-    .asStream().parallel().map {
-      fold("" to it) { (a, b), c ->
-        if (c == HOLE_MARKER) (a + " " + b.first()) to b.drop(1) else ("$a $c") to b
-      }.first.replace("ε ", "").trim()
-    }
-//      .filter { measureTimedValue { it.fastMatch(CFG) }.also { println("Decided ${it.value} in ${it.duration}") }.value }
-
 // This is fairly slow compared to bijective repair, but it's a good baseline for comparison
 fun repairInParallel(
   prompt: Σᐩ,