Skip to content

Latest commit

 

History

History
61 lines (44 loc) · 2.06 KB

File metadata and controls

61 lines (44 loc) · 2.06 KB

Roadmap

Active: Part 0 Reproduction and Evidence Map

  • Six-route literature and reproduction structure established.
  • Route 1 official LM-Polygraph estimator pilot completed.
  • Prompt-level stress labels separated from manually audited generation labels.

Next Route 1 milestones:

  1. Replace the diagnostic prompt set with claim-level factuality benchmarks.
  2. Add sampling-based semantic entropy and semantic-volume comparisons.
  3. Evaluate several text-only open-weight model families and decoding settings.
  4. Report calibration, compute cost, and risk--coverage with uncertainty intervals.

Completed: Part I Predictive Foundation

  • Bayesian regression benchmark.
  • Probabilistic scoring.
  • MCMC diagnostics.
  • Repeated-split comparison.
  • Part I reframed as the Bayesian predictive foundation for the doctoral programme.
  • README expanded into a linked multi-level research map.
  • Part III and Part IV documentation scaffolds.

Follow-up:

  • Audit Part I sampler and written prior consistency in a dedicated PR before any paper-style release.

Next Original-Method Study: Part II Bayesian Hallucination Risk Modeling

Milestones:

  1. Part II documentation scaffold.
  2. Text-only hallucination-risk dataset prototype.
  3. Feature extraction for uncertainty, consistency, and evidence.
  4. Bayesian logistic risk model.
  5. Calibration metrics: NLL, Brier, ECE, AUROC, AUPRC.
  6. Risk-coverage analysis.
  7. Explicit observation models for noisy labels and correlated evidence signals.

Planned: Part III Multimodal Hallucination Uncertainty

Milestones:

  1. Choose multimodal hallucination benchmarks.
  2. Define hallucination types.
  3. Extract visual grounding / evidence features.
  4. Fit hierarchical Bayesian risk models.
  5. Evaluate by hallucination type and dataset.

Planned: Part IV Bayesian Abstention and Decision Rules

Milestones:

  1. Define actions: answer, abstain, verify, regenerate.
  2. Define decision costs.
  3. Compare posterior-mean thresholds with credible-upper-bound thresholds.
  4. Evaluate selective risk and coverage.
  5. Study practical reliability trade-offs.