Reference · independent

RLCD vs RLHF

Both acronyms describe reinforcement learning, but they optimise for different things. RLHF — reinforcement learning from human feedback — is used to make generated text match what people prefer. RLCD — reinforcement learning for calibrated decisions — appears in the materials for Jev, Laya, and Clef as the objective behind their probability outputs.

The distinction matters because a decision API is only useful if its confidence number means something. If it does not, the threshold in your code is arbitrary.

Two different objectives

RLHF shapes a generative policy: which of two answers a human would rather read. The output is still free-form text, and the training signal is a preference over completions. It says nothing about whether a confidence number attached to a label is a frequency you can threshold.

RLCD shapes a scoring head over a fixed answer space. The training signal rewards probabilities that match observed outcomes — a proper scoring rule, with Brier loss named in Cloudflare's Clef materials. The point is that a 0.9 should be right about nine times in ten on comparable inputs, so code can act on it.

Side by side

DimensionRLHFRLCD
What is optimisedHuman preference over generated textCalibration of probabilities over a fixed answer set
OutputFree-form textTyped value with a probability per option
Typical signalPreference comparisons / reward modelProper scoring rules, Brier loss, calibration targets
Why it mattersAnswers read better and follow instructionsThresholds in code mean something measurable
Appears inGeneral-purpose chat and reasoning modelsDecision / System One model materials, including Jev, Laya, Clef

The comparison is about objectives, not a claim that either technique is exclusive to one model family. Vendor materials describe their own training; full recipes are rarely published.

Why calibration is the load-bearing claim

Of everything claimed about decision models, calibration is the one production systems actually depend on. Latency and price are easy to measure. A probability you threshold on is not — and the published evidence is thin. Laya reports mean expected calibration error improving from 0.466 to 0.081 after temperature fitting in its own fine-tune notebook. Clef claims Brier loss plus RLCD training.

An independent race cited on the Jev vs Clef comparison found a sharp split in practice: Jev scored around 0.99 confidence when right and 0.71 when wrong, while Clef sat near 0.81 when right and 0.74 when wrong. That is one test on one set of tasks, not an audit, but it is a good reason to measure calibration on your own labels before you automate anything.

What this means when you set a threshold

  • →Pick the threshold from a labelled sample of your own decisions, not from a vendor example.
  • →Track the outcome after the branch so you can recalibrate when the input distribution shifts.
  • →Treat low confidence as a route to a human or a slower model, not as a failure to hide.
  • →Re-baseline after a model version change — calibration behaviour is not guaranteed across builds.

Frequently asked questions

Is RLCD a replacement for RLHF?

+

No. They solve different problems. RLHF improves generated text against human preferences; RLCD improves the reliability of probabilities over a fixed answer space. A decision model has no prose to align, so preference alignment is not the objective it needs.

Why do decision models need calibration at all?

+

Because the useful output is a probability, not a label. If a 0.9 is right only half the time, an auto-approve threshold at 0.8 is a coin flip with extra steps. Calibration is what turns the number into something code can safely branch on.

Can I trust the confidence number from these models?

+

Treat it as a starting point. All calibration claims quoted on this site are vendor- or project-reported, and no independent audit is published here. One independent race found the two leading families calibrate quite differently. Measure on your own data.

Do all decision models use RLCD?

+

No. Jev, Laya, and Clef materials name RLCD or related proper-scoring objectives. AWS describes Strands Decider 2B primarily in terms of its pointer-head architecture and published training recipe. Check each project's documentation.

Is RLHF just RLCD with human labels?

+

No. Human labels can appear in either training pipeline, but the objective differs: preference ranking over generated answers versus calibrated probabilities over a declared answer set. Confusing the two leads people to expect a chat-model confidence score to behave like a decision-model one.

Keep reading

Sources

Reviewed 2026-11-13. Vendor- and project-reported figures are labelled as such; none were independently reproduced by this site.

Reference · reviewed 2026-11-13 · first published 2026-11-13