Abstract
Chain-of-thought supervision often forces models to compress intermediate reasoning into short, grader-friendly summaries. That compression improves scores on some benchmarks — and quietly destroys the traces auditors need. Legible Chains is a training and evaluation method for keeping intermediate reasoning faithful and human-auditable without collapsing the chain into a slogan.
The goal is not shorter reasoning. The goal is reasoning a careful reader can follow, challenge, and trust.
The collapse problem
When reward models prefer brevity, models learn to:
- Skip uncertain branches
- Replace derivations with conclusions
- Hide contradictions behind fluent prose
The result looks like reasoning. It is often a post-hoc narrative.
What we mean by legibility
A chain is legible when a reviewer can answer three questions without guessing:
- What claims were introduced, and when?
- Which steps depend on which earlier steps?
- Where did the model express uncertainty or refuse a leap?
Method
Legible Chains combines three pieces:
| Component | Role |
|---|---|
| Structured scratchpad | Forces explicit premises, steps, and checks |
| Faithfulness probes | Detects answers that do not follow from the written chain |
| Audit-time compression | Summaries are derived after grading, never substituted for the chain |
Structured scratchpad
Each step is typed. A minimal schema:
PREMISE: <claim grounded in the prompt or prior step>
INFER: <local deduction>
CHECK: <what would falsify this step>
COMMIT: <what is now treated as established>
Models may still write freely inside those slots. The slots prevent the chain from dissolving into undifferentiated prose.
Faithfulness probes
We intervene on intermediate commitments and measure whether the final answer moves. If the answer is invariant to corrupted premises, the chain was decorative.
def faithfulness_gap(model, prompt, chain, corrupt_step):
clean = model.answer(prompt, chain=chain)
dirty = model.answer(prompt, chain=corrupt(chain, corrupt_step))
return clean != dirty # True => that step was load-bearing
Results (sketch)
On a held-out audit set, Legible Chains raised human recovery of load-bearing steps while holding task accuracy roughly flat:
- +18pp recovery of critical premises by independent reviewers
- −4% median chain length vs. unconstrained CoT (not the objective)
- ≈0 change on primary task accuracy within confidence intervals
Limitations
- Typed steps add latency and token cost.
- Some domains (poetry, open brainstorming) resist rigid schemas.
- Faithfulness probes are necessary but not sufficient for truth.
Takeaway
Supervising reasoning should preserve the path, not only the destination. Collapse is an optimization failure dressed up as clarity.
Interpretability · 2026 · 06 · draft for markdown render testing