[
  [
    "kb/sources/geometry-of-knowledge-extending-diversity-boundaries-llms.ingest.md",
    1,
    "NoveltyBench: The method consistently outperforms baselines across generation budgets (10\u201330 samples), uncovering new semantic classes while maintaining high utility scores. At k=30, the approach achieves 16.65 distinct outputs versus 13.60 for G2 guidance."
  ],
  [
    "kb/sources/geometry-of-knowledge-extending-diversity-boundaries-llms.ingest.md",
    2,
    "AUT (Divergent Thinking): Latent-space exploration reaches Top-1 originality of 4.99 \u2014 approaching the practical upper bound of 5 \u2014 compared to 4.93 for G2 and 4.58 for multi-agent discussion."
  ],
  [
    "kb/sources/gsm-dc-llm-reasoning-distracted-irrelevant-context.ingest.md",
    5,
    "Step Accuracy (SAcc): Our symbolic parser reads the model's chain-of-thought and executes every intermediate equation in topological order. SAcc = 1 iff *all* equations are arithmetically correct *and* each step references only symbols that have already been defined."
  ],
  [
    "kb/sources/gsm-dc-llm-reasoning-distracted-irrelevant-context.ingest.md",
    6,
    "Path Accuracy (PAcc): To quantify *distraction robustness* we check whether the model confines its reasoning to the augmented dependency graph G' after injecting irrelevant context. PAcc = 1 iff (i) every required dependency on P is present, and (ii) no irrelevant node is used in the computation of any step on P."
  ],
  [
    "kb/sources/meta-harness-end-to-end-optimization-of-model-harnesses.ingest.md",
    4,
    "| **Meta-Harness** (full) | \u2713 | \u2713 | - | \u2713 | 50.0 | 56.7 | 39 |"
  ],
  [
    "kb/sources/meyerson-maker-million-step-llm-zero-errors.ingest.md",
    3,
    "This paper provides a first implementation of the MDAP framework: MAKER (for Maximal Agentic decomposition, first-to-ahead-by-K Error correction, and Red-flagging), evaluated on the Towers of Hanoi domain. MAKER is a system of *agents* in which each agent is assigned a single subtask to solve."
  ],
  [
    "kb/sources/meyerson-maker-million-step-llm-zero-errors.ingest.md",
    4,
    "Maximal agentic decomposition (MAD), i.e., m = 1"
  ],
  [
    "kb/sources/paulsen-maximum-effective-context-window-mecw.ingest.md",
    4,
    "Dataset: 10,000 unique names of individuals, each assigned a random number 1-20, a random item from a list of 15 possibilities, and a random color out of 9 possibilities. Example data row: \"Abigail Holmes has 19 red balloons.\""
  ],
  [
    "kb/sources/paulsen-maximum-effective-context-window-mecw.ingest.md",
    5,
    "Question types (four distinct types): 1. Needle-in-a-Haystack: Search for a single data point \u2014 \"How many objects does {person} have?\" 2. Needles-in-a-Haystack: Search for multiple data points and sum \u2014 \"How many {color} objects are there?\" or \"How many {object} are there?\" 3. Summarization: Full sum of all data points \u2014 \"How many objects are there total?\" 4. Find and Sort: Search for multiple data points then sort alphabetically by name \u2014 \"Find all people with {color} objects. Sort them by first and last name. Concatenate the number of objects they have into one long string value in the order they were sorted.\""
  ],
  [
    "kb/sources/towards-faithfully-interpretable-nlp-systems.ingest.md",
    2,
    "Faithfulness evaluation should not involve human-judgement on the quality of interpretation. We note that: (1) humans cannot judge if an interpretation is faithful or not: if they understood the model, interpretation would be unnecessary; (2) for similar reasons, we cannot obtain supervision for this problem, either. Therefore, human judgement should not be involved in evaluation for faithfulness, as human judgement measures plausability."
  ]
]