When an AI rewrites its own draft, should it fix whatever breaks the rules, or polish the whole thing evenly?
Do draft-and-revise loops work better when guided by unresolved constraints than by diffusion-style denoising?
This explores whether a revise-the-draft loop should pick what to fix by checking which rules the draft still breaks, or by repeatedly smoothing the whole draft the way a diffusion model removes noise.
This explores whether a revision loop should decide what to fix by asking which rules the draft still breaks, or by smoothing the whole draft step by step, the way diffusion models remove noise. The corpus has no head-to-head test of the two. Read together, though, the notes suggest they do different jobs. Diffusion describes how revision is paced. Constraints supply the ability to say "this part is wrong."
The case for the diffusion framing is in Can iterative revision cycles match how humans actually write?. It treats a research report as a persistent draft skeleton that counts as noisy output and gets cleaned up through targeted retrieval. That keeps the whole report coherent better than a linear pipeline, and it matches how people write. But the comparison is against linear pipelines, not against constraint-guided loops. The "noise" it removes is mostly thin or shaky content that more evidence can fix.
The constraint case starts from a limit of the architecture. Why does autoregressive generation fail at constraint satisfaction? argues that autoregressive models can't take back a token once it's emitted. Constraint solvers work by discarding partial answers that break a rule, and adding a symbolic solver helps because it supplies that missing move. A draft-and-revise loop is one way to give a text generator retraction back. Once you have it, the question is what triggers it. A specific unresolved constraint is a concrete trigger. A generic "make it cleaner" pass is not. [[llms-cannot-execute-iterative-numerical-methods-in-latent-space-and-fall-back-to]} points the same way. Models don't run iterative procedures internally. They pattern-match to a plausible answer. So the refining has to happen in visible drafts that something outside the model checks. The loop around the model does the denoising, not the model itself.
There's a reason denoising may be the weaker guide when rules are involved. Can models trained on many imperfect experts outperform everyone? shows that consensus and low-temperature sampling wash out uncorrelated errors. That is the kind of mistake averaging can fix. A broken requirement is not random. It's systematic, and pulling toward the most typical output can keep it in place. This is my inference from the notes, not something either note tests. Two other notes hint at what a good guiding signal looks like. Does step-level confidence outperform global averaging for trace filtering? finds that local step-level checks catch breakdowns that a global average hides. That is the same logic as aiming a revision at the one unmet constraint rather than at the draft's overall quality. Can verification separate structural near-misses from topical matches? shows that a verifier looking at full token-to-token patterns rejects near-misses that compressed similarity scores accept. A draft can look on-topic and still break a structural requirement, so the checker has to be able to see that.
The tentative answer is that constraints should decide what gets revised, and a coarse-to-fine, diffusion-like schedule can still decide the order and keep the piece coherent. Do critique models improve diversity during training itself? adds a caution from training. Step-level critique that names a specific problem keeps solutions diverse and stops premature convergence. That suggests targeted feedback is safer than pulling everything toward a consensus draft. The untested gap is open-ended writing, where nobody has written down the constraints. There the diffusion-style loop may be all that's available.
Sources 7 notes
Research writing follows a draft-and-revise pattern analogous to diffusion sampling, where a persistent draft skeleton is iteratively denoised through targeted retrieval steps. This architecture maintains global coherence better than linear pipelines while mirroring cognitive studies of actual human writing.
The performance ceiling on constraint satisfaction problems is not a model-quality issue but an architectural limitation: autoregressive transformers cannot retract emitted tokens, while CSP solvers fundamentally depend on discarding invalid partial assignments. Symbolic solver integration works because it supplies what the architecture lacks.
Research shows LLMs cannot perform iterative procedures in latent space. They recognize optimization problems as template-similar and emit plausible-looking but incorrect values, a failure mode that persists across model scale and training approaches.
Models trained on diverse experts converge on consensus behavior that outperforms individuals. Low-temperature sampling concentrates outputs on this majority-voted consensus, denoising uncorrelated biases and errors across the training set.
Local step-level confidence catches reasoning breakdowns that global averaging masks and enables early stopping before traces complete. This approach achieves comparable accuracy gains to naive majority voting with far fewer generated traces, proving trace quality matters more than quantity.
Show all 7 sources
A two-stage pipeline—pooled-cosine recall followed by a small Transformer verifier operating on token-token similarity maps—reliably rejects structural near-misses that MaxSim-style late interaction cannot. The verifier succeeds because it operates on full token interaction patterns rather than compressed vectors.
Step-level critique in the training loop counteracts tail narrowing and maintains solution diversity across self-training iterations. This training-time benefit—preventing premature convergence—is more fundamental than test-time accuracy gains.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Can Large Language Models Reason and Optimize Under Constraints?
- Deep Think with Confidence
- Training for Compositional Sensitivity Reduces Dense Retrieval Generalization
- Deep Researcher with Test-Time Diffusion
- Transcendence: Generative Models Can Outperform The Experts That Train Them
- Branch-Solve-Merge Improves Large Language Model Evaluation and Generation
- Large Language Diffusion Models
- Local Coherence or Global Validity? Investigating RLVR Traces in Math Domains