Lesson overview | Previous part | Next part

Preference Optimization RLHF and DPO: Part 3: Reward Modeling

3. Reward Modeling

Reward Modeling develops the part of preference optimization rlhf and dpo that the approved TOC assigns to Chapter 18. The emphasis is alignment behavior, safety constraints, and feedback loops, not generic fine-tuning or production monitoring.

3.1 Pairwise logistic loss

Pairwise logistic loss belongs in the canonical scope of preference optimization rlhf and dpo. The object is the preference-aligned policy, not merely a prompt trick or a moderation label. We study how data, losses, policies, review processes, and safety constraints shape a model's conditional distribution over responses.

A compact way to read this subsection is through the local symbol (x,y_w,y_l). It marks the alignment object being transformed: an instruction policy, a preference pair, a violation classifier, a guardrail action, or a feedback event. The details differ, but the discipline is the same: state the object, state the loss or decision rule, then audit the behavioral side effects.

$$\mathcal{L}_{\mathrm{DPO}}(\theta) = -\mathbb{E}\log\sigma\left(\beta\log\frac{\pi_\theta(y_w\mid x)}{\pi_{\mathrm{ref}}(y_w\mid x)}-\beta\log\frac{\pi_\theta(y_l\mid x)}{\pi_{\mathrm{ref}}(y_l\mid x)}\right).$$

For pairwise logistic loss, this formula should not be treated as a slogan. It defines which tokens, responses, comparisons, or decisions receive gradient or operational weight. A change in masking, sampling, rubric wording, or thresholding changes the effective objective even if the model architecture is unchanged.

Examples:

• Treat pairwise logistic loss as part of the model contract and store the exact data version.
• Record the prompt template, role format, policy version, and decoder settings.
• Compare aligned and reference policies on both helpfulness and safety slices.
• Use held-out examples that were not used to tune refusals or rewards.
• Inspect failure cases before declaring the objective successful.

Non-examples:

• Calling a model aligned because it sounds polite on a few prompts.
• Training on refusals without measuring over-refusal on benign requests.
• Using a reward model as ground truth without calibration or adversarial checks.
• Shipping a guardrail threshold without measuring false positive and false negative rates.
• Letting feedback logs change training without provenance or consent controls.

A useful implementation pattern is to separate policy, data, and measurement. The policy says what behavior is desired. The data supplies examples, comparisons, attacks, or feedback events. The measurement checks whether the updated system moved in the intended direction without unacceptable regressions.

policy text/rubric
      |
      v
training or guardrail data  ->  objective/threshold  ->  aligned system
      |                                                   |
      v                                                   v
audit metadata                                      held-out safety eval

Worked reasoning pattern for pairwise logistic loss:

1. Name the target behavior in plain language.
2. Write the mathematical variable that represents it.
3. Specify which examples or comparisons estimate it.
4. Choose the optimization loss or runtime decision rule.
5. Define the regression metric that would prove the change became worse.

Three details are especially easy to miss in alignment work. First, the user intent distribution is not the same as the pretraining distribution. Second, safety labels are not ordinary class labels; they encode policy judgments that can change by context. Third, optimization pressure finds shortcuts, so every proxy must be monitored for Goodhart-style failures.

AI connection: Pairwise logistic loss is part of the post-training stack used by modern assistant systems. It links the base language model to human intent, safety policy, and deployment constraints without pretending that a single loss can capture all values. The goal is not perfect alignment by formula; it is a repeatable loop where evidence, objectives, and safeguards improve together.

3.2 Reward margins

Reward margins belongs in the canonical scope of preference optimization rlhf and dpo. The object is the preference-aligned policy, not merely a prompt trick or a moderation label. We study how data, losses, policies, review processes, and safety constraints shape a model's conditional distribution over responses.

A compact way to read this subsection is through the local symbol (x,y_w,y_l). It marks the alignment object being transformed: an instruction policy, a preference pair, a violation classifier, a guardrail action, or a feedback event. The details differ, but the discipline is the same: state the object, state the loss or decision rule, then audit the behavioral side effects.

$$\mathcal{L}_{\mathrm{DPO}}(\theta) = -\mathbb{E}\log\sigma\left(\beta\log\frac{\pi_\theta(y_w\mid x)}{\pi_{\mathrm{ref}}(y_w\mid x)}-\beta\log\frac{\pi_\theta(y_l\mid x)}{\pi_{\mathrm{ref}}(y_l\mid x)}\right).$$

For reward margins, this formula should not be treated as a slogan. It defines which tokens, responses, comparisons, or decisions receive gradient or operational weight. A change in masking, sampling, rubric wording, or thresholding changes the effective objective even if the model architecture is unchanged.

Examples:

• Treat reward margins as part of the model contract and store the exact data version.
• Record the prompt template, role format, policy version, and decoder settings.
• Compare aligned and reference policies on both helpfulness and safety slices.
• Use held-out examples that were not used to tune refusals or rewards.
• Inspect failure cases before declaring the objective successful.

Non-examples:

• Calling a model aligned because it sounds polite on a few prompts.
• Training on refusals without measuring over-refusal on benign requests.
• Using a reward model as ground truth without calibration or adversarial checks.
• Shipping a guardrail threshold without measuring false positive and false negative rates.
• Letting feedback logs change training without provenance or consent controls.

A useful implementation pattern is to separate policy, data, and measurement. The policy says what behavior is desired. The data supplies examples, comparisons, attacks, or feedback events. The measurement checks whether the updated system moved in the intended direction without unacceptable regressions.

policy text/rubric
      |
      v
training or guardrail data  ->  objective/threshold  ->  aligned system
      |                                                   |
      v                                                   v
audit metadata                                      held-out safety eval

Worked reasoning pattern for reward margins:

1. Name the target behavior in plain language.
2. Write the mathematical variable that represents it.
3. Specify which examples or comparisons estimate it.
4. Choose the optimization loss or runtime decision rule.
5. Define the regression metric that would prove the change became worse.

Three details are especially easy to miss in alignment work. First, the user intent distribution is not the same as the pretraining distribution. Second, safety labels are not ordinary class labels; they encode policy judgments that can change by context. Third, optimization pressure finds shortcuts, so every proxy must be monitored for Goodhart-style failures.

AI connection: Reward margins is part of the post-training stack used by modern assistant systems. It links the base language model to human intent, safety policy, and deployment constraints without pretending that a single loss can capture all values. The goal is not perfect alignment by formula; it is a repeatable loop where evidence, objectives, and safeguards improve together.

3.3 Annotator disagreement

Annotator disagreement belongs in the canonical scope of preference optimization rlhf and dpo. The object is the preference-aligned policy, not merely a prompt trick or a moderation label. We study how data, losses, policies, review processes, and safety constraints shape a model's conditional distribution over responses.

A compact way to read this subsection is through the local symbol (x,y_w,y_l). It marks the alignment object being transformed: an instruction policy, a preference pair, a violation classifier, a guardrail action, or a feedback event. The details differ, but the discipline is the same: state the object, state the loss or decision rule, then audit the behavioral side effects.

$$\mathcal{L}_{\mathrm{DPO}}(\theta) = -\mathbb{E}\log\sigma\left(\beta\log\frac{\pi_\theta(y_w\mid x)}{\pi_{\mathrm{ref}}(y_w\mid x)}-\beta\log\frac{\pi_\theta(y_l\mid x)}{\pi_{\mathrm{ref}}(y_l\mid x)}\right).$$

For annotator disagreement, this formula should not be treated as a slogan. It defines which tokens, responses, comparisons, or decisions receive gradient or operational weight. A change in masking, sampling, rubric wording, or thresholding changes the effective objective even if the model architecture is unchanged.

Examples:

• Treat annotator disagreement as part of the model contract and store the exact data version.
• Record the prompt template, role format, policy version, and decoder settings.
• Compare aligned and reference policies on both helpfulness and safety slices.
• Use held-out examples that were not used to tune refusals or rewards.
• Inspect failure cases before declaring the objective successful.

Non-examples:

• Calling a model aligned because it sounds polite on a few prompts.
• Training on refusals without measuring over-refusal on benign requests.
• Using a reward model as ground truth without calibration or adversarial checks.
• Shipping a guardrail threshold without measuring false positive and false negative rates.
• Letting feedback logs change training without provenance or consent controls.

A useful implementation pattern is to separate policy, data, and measurement. The policy says what behavior is desired. The data supplies examples, comparisons, attacks, or feedback events. The measurement checks whether the updated system moved in the intended direction without unacceptable regressions.

policy text/rubric
      |
      v
training or guardrail data  ->  objective/threshold  ->  aligned system
      |                                                   |
      v                                                   v
audit metadata                                      held-out safety eval

Worked reasoning pattern for annotator disagreement:

1. Name the target behavior in plain language.
2. Write the mathematical variable that represents it.
3. Specify which examples or comparisons estimate it.
4. Choose the optimization loss or runtime decision rule.
5. Define the regression metric that would prove the change became worse.

Three details are especially easy to miss in alignment work. First, the user intent distribution is not the same as the pretraining distribution. Second, safety labels are not ordinary class labels; they encode policy judgments that can change by context. Third, optimization pressure finds shortcuts, so every proxy must be monitored for Goodhart-style failures.

AI connection: Annotator disagreement is part of the post-training stack used by modern assistant systems. It links the base language model to human intent, safety policy, and deployment constraints without pretending that a single loss can capture all values. The goal is not perfect alignment by formula; it is a repeatable loop where evidence, objectives, and safeguards improve together.

3.4 Reward calibration

Reward calibration belongs in the canonical scope of preference optimization rlhf and dpo. The object is the preference-aligned policy, not merely a prompt trick or a moderation label. We study how data, losses, policies, review processes, and safety constraints shape a model's conditional distribution over responses.

A compact way to read this subsection is through the local symbol (x,y_w,y_l). It marks the alignment object being transformed: an instruction policy, a preference pair, a violation classifier, a guardrail action, or a feedback event. The details differ, but the discipline is the same: state the object, state the loss or decision rule, then audit the behavioral side effects.

$$\mathcal{L}_{\mathrm{DPO}}(\theta) = -\mathbb{E}\log\sigma\left(\beta\log\frac{\pi_\theta(y_w\mid x)}{\pi_{\mathrm{ref}}(y_w\mid x)}-\beta\log\frac{\pi_\theta(y_l\mid x)}{\pi_{\mathrm{ref}}(y_l\mid x)}\right).$$

For reward calibration, this formula should not be treated as a slogan. It defines which tokens, responses, comparisons, or decisions receive gradient or operational weight. A change in masking, sampling, rubric wording, or thresholding changes the effective objective even if the model architecture is unchanged.

Examples:

• Treat reward calibration as part of the model contract and store the exact data version.
• Record the prompt template, role format, policy version, and decoder settings.
• Compare aligned and reference policies on both helpfulness and safety slices.
• Use held-out examples that were not used to tune refusals or rewards.
• Inspect failure cases before declaring the objective successful.

Non-examples:

• Calling a model aligned because it sounds polite on a few prompts.
• Training on refusals without measuring over-refusal on benign requests.
• Using a reward model as ground truth without calibration or adversarial checks.
• Shipping a guardrail threshold without measuring false positive and false negative rates.
• Letting feedback logs change training without provenance or consent controls.

A useful implementation pattern is to separate policy, data, and measurement. The policy says what behavior is desired. The data supplies examples, comparisons, attacks, or feedback events. The measurement checks whether the updated system moved in the intended direction without unacceptable regressions.

policy text/rubric
      |
      v
training or guardrail data  ->  objective/threshold  ->  aligned system
      |                                                   |
      v                                                   v
audit metadata                                      held-out safety eval

Worked reasoning pattern for reward calibration:

1. Name the target behavior in plain language.
2. Write the mathematical variable that represents it.
3. Specify which examples or comparisons estimate it.
4. Choose the optimization loss or runtime decision rule.
5. Define the regression metric that would prove the change became worse.

Three details are especially easy to miss in alignment work. First, the user intent distribution is not the same as the pretraining distribution. Second, safety labels are not ordinary class labels; they encode policy judgments that can change by context. Third, optimization pressure finds shortcuts, so every proxy must be monitored for Goodhart-style failures.

AI connection: Reward calibration is part of the post-training stack used by modern assistant systems. It links the base language model to human intent, safety policy, and deployment constraints without pretending that a single loss can capture all values. The goal is not perfect alignment by formula; it is a repeatable loop where evidence, objectives, and safeguards improve together.

3.5 RewardBench-style evaluation

RewardBench-style evaluation belongs in the canonical scope of preference optimization rlhf and dpo. The object is the preference-aligned policy, not merely a prompt trick or a moderation label. We study how data, losses, policies, review processes, and safety constraints shape a model's conditional distribution over responses.

A compact way to read this subsection is through the local symbol (x,y_w,y_l). It marks the alignment object being transformed: an instruction policy, a preference pair, a violation classifier, a guardrail action, or a feedback event. The details differ, but the discipline is the same: state the object, state the loss or decision rule, then audit the behavioral side effects.

$$\mathcal{L}_{\mathrm{DPO}}(\theta) = -\mathbb{E}\log\sigma\left(\beta\log\frac{\pi_\theta(y_w\mid x)}{\pi_{\mathrm{ref}}(y_w\mid x)}-\beta\log\frac{\pi_\theta(y_l\mid x)}{\pi_{\mathrm{ref}}(y_l\mid x)}\right).$$

For rewardbench-style evaluation, this formula should not be treated as a slogan. It defines which tokens, responses, comparisons, or decisions receive gradient or operational weight. A change in masking, sampling, rubric wording, or thresholding changes the effective objective even if the model architecture is unchanged.

Examples:

• Treat rewardbench-style evaluation as part of the model contract and store the exact data version.
• Record the prompt template, role format, policy version, and decoder settings.
• Compare aligned and reference policies on both helpfulness and safety slices.
• Use held-out examples that were not used to tune refusals or rewards.
• Inspect failure cases before declaring the objective successful.

Non-examples:

• Calling a model aligned because it sounds polite on a few prompts.
• Training on refusals without measuring over-refusal on benign requests.
• Using a reward model as ground truth without calibration or adversarial checks.
• Shipping a guardrail threshold without measuring false positive and false negative rates.
• Letting feedback logs change training without provenance or consent controls.

A useful implementation pattern is to separate policy, data, and measurement. The policy says what behavior is desired. The data supplies examples, comparisons, attacks, or feedback events. The measurement checks whether the updated system moved in the intended direction without unacceptable regressions.

policy text/rubric
      |
      v
training or guardrail data  ->  objective/threshold  ->  aligned system
      |                                                   |
      v                                                   v
audit metadata                                      held-out safety eval

Worked reasoning pattern for rewardbench-style evaluation:

1. Name the target behavior in plain language.
2. Write the mathematical variable that represents it.
3. Specify which examples or comparisons estimate it.
4. Choose the optimization loss or runtime decision rule.
5. Define the regression metric that would prove the change became worse.

Three details are especially easy to miss in alignment work. First, the user intent distribution is not the same as the pretraining distribution. Second, safety labels are not ordinary class labels; they encode policy judgments that can change by context. Third, optimization pressure finds shortcuts, so every proxy must be monitored for Goodhart-style failures.

AI connection: RewardBench-style evaluation is part of the post-training stack used by modern assistant systems. It links the base language model to human intent, safety policy, and deployment constraints without pretending that a single loss can capture all values. The goal is not perfect alignment by formula; it is a repeatable loop where evidence, objectives, and safeguards improve together.