Speech & Language Processing Chapter 9 — Post-training: Instruction Tuning, Alignment, and Test-Time Compute Explained

// the one-minute version

Pretraining teaches a model to continue text; post-training teaches it to follow instructions and behave according to selected preferences. Supervised instruction tuning learns from prompt–response examples. Preference learning uses comparisons between candidate responses and may train a reward model followed by reinforcement learning, or optimize preferences more directly with methods such as DPO. Test-time compute improves some answers by generating, critiquing, searching, or verifying additional candidates. None of these steps creates a universal human value function: data collection, annotator variation, reward hacking, over-refusal, and evaluation design remain central.

Imagine arriving in a research lab with one deceptively simple request: build a system that can use language reliably. The room already contains datasets, model checkpoints, annotation manuals, benchmark tables, and confident demos. What it does not contain is a guarantee that everybody means the same thing by “use language.” Chapter 9 slows the room down. Its subject is Post-training: Instruction Tuning, Alignment, and Test-Time Compute. Its job is to turn a broad topic into a sequence of explicit choices that a researcher can inspect, test, and defend.

Our running story will not begin with a library call. It begins with a user and an observable signal. Begin with supervised examples that show concise diagnosis, evidence requests, and safe commands. Collect pairwise preferences where reviewers choose grounded, reversible guidance over guesses or destructive actions. Optimize the model while monitoring drift from its base capabilities. At inference time, generate two plans, run a policy and command-risk checker, and select the plan whose claims are supported by telemetry. Evaluate helpfulness and refusal separately so the model cannot “win” safety by refusing everything. That miniature situation is enough to expose the chapter's full problem. Something in the world becomes data; the data is represented; alternatives are scored; a decision is decoded; and somebody must decide whether the result is good enough to use. The model is only one character in that story.

This companion follows the structure and main ideas of the official January 6, 2026 draft, but the wording, examples, diagrams, exercises, and explanations are original. It is written for a master's student who needs more than a list of terms: you should be able to derive the central mechanism, identify its assumptions, design a controlled experiment, and explain why a strong benchmark number may still fail to establish the claim you care about.

chapter promiseBy the end, you should be able to explain post-training: instruction tuning, alignment, and test-time compute from first principles, connect its major components, work through a concrete case, recognize the most common invalid shortcuts, and design an evaluation that separates model quality from data leakage or measurement error.

01 Begin with the problem, not the model

Supervised instruction tuning, preference data, reward modeling, RLHF-style optimization, direct preference optimization, alignment limits, reasoning, and test-time compute. The important noun in that sentence is not the name of an architecture. It is the problem. A problem statement identifies an observable input, a desired output, the context legitimately available at decision time, and the cost of being wrong. Architecture selection comes later.

First principles are useful because language systems accumulate invisible conventions. A word boundary may be assumed rather than defined. A label may mix several human judgments. A train/test split may let the same speaker, document template, or memorized passage appear on both sides. An aggregate metric may give every error equal value even when deployment does not. If those choices remain hidden, a larger model can improve the number while leaving the real problem untouched.

For this chapter, ask four questions before reading any result. Representation: what exactly becomes a token, vector, frame, span, state, or candidate? Objective: which quantity is optimized, and is it the same behavior users need? Inference: how are candidates searched, constrained, ranked, or sampled? Evaluation: what held-out evidence would falsify the claim? These four questions create a map that survives changes in implementation fashion.

The loop used throughout this companion: make each hidden choice inspectable before trusting the final behavior.

02 The chapter's conceptual map

Read the following concepts as cooperating modules rather than vocabulary for an exam. Each owns a distinct responsibility. The interfaces between them are where assumptions become visible and where most serious debugging begins.

1. Instruction tuning changes the interface

A curated dataset pairs diverse instructions with desired responses. Fine-tuning makes following the instruction more likely than merely continuing its surface form. Coverage, response quality, and formatting diversity matter more than raw example count alone.

Build this idea from its contract. First name the observation available to the system; then name the representation that preserves the useful part of that observation; then identify the score, probability, constraint, or rule that distinguishes one candidate from another. Finally ask what output leaves this component and which later component is allowed to trust it. In post-training: instruction tuning, alignment, and test-time compute, these boundaries matter because a superficially correct final answer can be produced by the wrong evidence. A master's-level analysis therefore explains not just what the component returns, but what information it discards and which assumptions make the return value meaningful.

The idea also has an operational side. We would want examples where instruction tuning changes the interface clearly helps, counterexamples where its assumptions break, an ablation that removes it, and a metric sensitive to the expected change. If removing it changes nothing, either the rest of the system learned a substitute or the evaluation never exercised the behavior. If it improves the average but harms a language, subgroup, input length, or rare class, the average is incomplete evidence. This is how we move from remembering a definition to making and defending a research claim.

Keep instruction tuning changes the interface separate from preferences are relative labels. They cooperate, but they answer different questions. Collapsing the two makes debugging impossible: an engineer sees only an incorrect output and cannot tell whether the representation was weak, the candidate set was incomplete, the scoring function preferred the wrong item, or the decoder violated a constraint. The chapter's larger lesson is to preserve these interfaces even when a neural model learns several stages jointly.

2. Preferences are relative labels

Annotators compare two or more responses under a rubric. A reward model can learn to score candidates, but it learns the dataset’s judgments and shortcuts. Agreement, annotator context, and prompt distribution should be documented.

The idea also has an operational side. We would want examples where preferences are relative labels clearly helps, counterexamples where its assumptions break, an ablation that removes it, and a metric sensitive to the expected change. If removing it changes nothing, either the rest of the system learned a substitute or the evaluation never exercised the behavior. If it improves the average but harms a language, subgroup, input length, or rare class, the average is incomplete evidence. This is how we move from remembering a definition to making and defending a research claim.

Keep preferences are relative labels separate from rlhf and direct preference methods. They cooperate, but they answer different questions. Collapsing the two makes debugging impossible: an engineer sees only an incorrect output and cannot tell whether the representation was weak, the candidate set was incomplete, the scoring function preferred the wrong item, or the decoder violated a constraint. The chapter's larger lesson is to preserve these interfaces even when a neural model learns several stages jointly.

3. RLHF and direct preference methods

RLHF commonly optimizes a policy against a learned reward while constraining drift from a reference model. Direct preference optimization reframes preference pairs into a classification-like objective without a separate online reinforcement-learning loop. Both depend on the quality and scope of comparisons.

The idea also has an operational side. We would want examples where rlhf and direct preference methods clearly helps, counterexamples where its assumptions break, an ablation that removes it, and a metric sensitive to the expected change. If removing it changes nothing, either the rest of the system learned a substitute or the evaluation never exercised the behavior. If it improves the average but harms a language, subgroup, input length, or rare class, the average is incomplete evidence. This is how we move from remembering a definition to making and defending a research claim.

Keep rlhf and direct preference methods separate from alignment has competing objectives. They cooperate, but they answer different questions. Collapsing the two makes debugging impossible: an engineer sees only an incorrect output and cannot tell whether the representation was weak, the candidate set was incomplete, the scoring function preferred the wrong item, or the decoder violated a constraint. The chapter's larger lesson is to preserve these interfaces even when a neural model learns several stages jointly.

4. Alignment has competing objectives

Helpfulness, harmlessness, honesty, style, and user autonomy can conflict. A single scalar reward compresses those tensions. Models may over-refuse safe requests, flatter users, exploit reward-model weaknesses, or perform well on familiar safety tests while failing under distribution shift.

The idea also has an operational side. We would want examples where alignment has competing objectives clearly helps, counterexamples where its assumptions break, an ablation that removes it, and a metric sensitive to the expected change. If removing it changes nothing, either the rest of the system learned a substitute or the evaluation never exercised the behavior. If it improves the average but harms a language, subgroup, input length, or rare class, the average is incomplete evidence. This is how we move from remembering a definition to making and defending a research claim.

Keep alignment has competing objectives separate from test-time compute is an inference budget. They cooperate, but they answer different questions. Collapsing the two makes debugging impossible: an engineer sees only an incorrect output and cannot tell whether the representation was weak, the candidate set was incomplete, the scoring function preferred the wrong item, or the decoder violated a constraint. The chapter's larger lesson is to preserve these interfaces even when a neural model learns several stages jointly.

5. Test-time compute is an inference budget

A system can sample multiple solutions, ask for critique, search a reasoning tree, call tools, or run a verifier. More compute helps only when candidate diversity and selection correlate with correctness; otherwise it multiplies confident errors and cost.

The idea also has an operational side. We would want examples where test-time compute is an inference budget clearly helps, counterexamples where its assumptions break, an ablation that removes it, and a metric sensitive to the expected change. If removing it changes nothing, either the rest of the system learned a substitute or the evaluation never exercised the behavior. If it improves the average but harms a language, subgroup, input length, or rare class, the average is incomplete evidence. This is how we move from remembering a definition to making and defending a research claim.

Keep test-time compute is an inference budget separate from instruction tuning changes the interface. They cooperate, but they answer different questions. Collapsing the two makes debugging impossible: an engineer sees only an incorrect output and cannot tell whether the representation was weak, the candidate set was incomplete, the scoring function preferred the wrong item, or the decoder violated a constraint. The chapter's larger lesson is to preserve these interfaces even when a neural model learns several stages jointly.

compressionThe chapter can be remembered as a chain: define the unit, preserve the relevant evidence, score alternatives under explicit assumptions, decode a legal result, then evaluate the behavior at the same grain as the real decision.

03 Derive the core mechanism carefully

Theory is useful when every symbol has an operational meaning. Do not memorize an equation before identifying what produces each term, which terms are observed, which are learned, and which approximation makes computation possible.

// the central relationshipPreference optimization

maximize log σ(β[(log πθ(y⁺|x)−log πref(y⁺|x))−(log πθ(y⁻|x)−log πref(y⁻|x))])A preferred response y⁺ should become more likely than a rejected response y⁻ relative to a reference policy. The temperature-like β controls how strongly the model moves. This objective can simplify optimization, but it cannot repair ambiguous rubrics, unrepresentative prompts, annotator disagreement, or missing safety cases. The data defines the direction of alignment.

Now derive the system around the relationship. Start with the smallest legal input and write down its shape or structure. Enumerate the candidate outputs. Calculate or reason through one score by hand. Check normalization or structural constraints. Then change one input feature and predict the direction of the output change before executing code. This procedure catches sign errors, leaked context, illegal transitions, and confused units far earlier than end-to-end benchmarking.

Next distinguish estimation from decision. A model may estimate probabilities, similarities, alignments, or scores. A decoder, threshold, search algorithm, or policy turns them into a discrete output. The best estimator under log loss may not produce the best operational decision under asymmetric costs. Conversely, a clever decoder can hide a weak model on one benchmark while failing when the candidate distribution changes.

Finally distinguish training objective from evaluation metric. We choose differentiable losses because gradient-based optimization needs them; we choose evaluation measures because people need evidence about behavior. The two should be related but need not be identical. When they diverge, state the reason and test whether improvement in the surrogate actually predicts improvement in the target behavior.

04 A worked story from input to evidence

Training a safer troubleshooting assistant

Begin with supervised examples that show concise diagnosis, evidence requests, and safe commands. Collect pairwise preferences where reviewers choose grounded, reversible guidance over guesses or destructive actions. Optimize the model while monitoring drift from its base capabilities. At inference time, generate two plans, run a policy and command-risk checker, and select the plan whose claims are supported by telemetry. Evaluate helpfulness and refusal separately so the model cannot “win” safety by refusing everything.

Stage 1 — define the observation. Record what the system truly receives at decision time. Do not quietly add future text, a gold annotation, a clean transcript, a manually selected passage, or metadata unavailable in production. This stage protects the validity of everything after it.

Stage 2 — construct the representation. Choose units that preserve the distinctions the task needs while remaining learnable from available data. Document normalization, vocabulary, missing values, masking, and alignment. Save enough information to map predictions back to the original input.

Stage 3 — produce candidates and scores. The model turns evidence into alternatives. Inspect at least the winner, a plausible runner-up, and an obviously wrong candidate. Their score differences reveal whether the model has a robust preference or won by a tiny, unstable margin.

Stage 4 — apply constraints and policy. A legal sequence, supported citation, safe action, or valid structure may require rules beyond the learned score. This is also where uncertainty becomes an abstention, clarification, escalation, or request for more evidence rather than a forced guess.

Stage 5 — evaluate at several levels. Component metrics localize faults; end-to-end metrics measure user-visible behavior. Use both. A correct final result can conceal a broken intermediate stage, and a strong component can be neutralized by a bad downstream policy.

Stage 6 — perform error analysis. Group failures by mechanism instead of collecting anecdotes. Look for length, frequency, language, subgroup, domain, noise, ambiguity, and annotation effects. A model improvement becomes scientifically convincing when it fixes the predicted category without creating an unreported regression elsewhere.

05 What usually goes wrong

The following traps are not footnotes. They are common ways a technically correct implementation produces a misleading research conclusion or unsafe product behavior.

common catches & gotchas

Treating annotator preference as objective ground truth. The visible symptom is a result that may look plausible on ordinary examples while failing when this assumption is stressed. The likely cause is that training or evaluation rewarded a shortcut. Correct it by adding a targeted counterexample, measuring this failure separately, and tracing the decision back to its source evidence before changing model size.
Optimizing a reward score without checking for reward hacking or style shortcuts. The visible symptom is a result that may look plausible on ordinary examples while failing when this assumption is stressed. The likely cause is that training or evaluation rewarded a shortcut. Correct it by adding a targeted counterexample, measuring this failure separately, and tracing the decision back to its source evidence before changing model size.
Measuring safety only by refusal rate. The visible symptom is a result that may look plausible on ordinary examples while failing when this assumption is stressed. The likely cause is that training or evaluation rewarded a shortcut. Correct it by adding a targeted counterexample, measuring this failure separately, and tracing the decision back to its source evidence before changing model size.
Using more reasoning tokens without a reliable verifier. The visible symptom is a result that may look plausible on ordinary examples while failing when this assumption is stressed. The likely cause is that training or evaluation rewarded a shortcut. Correct it by adding a targeted counterexample, measuring this failure separately, and tracing the decision back to its source evidence before changing model size.
Ignoring preference differences across cultures, domains, and user roles. The visible symptom is a result that may look plausible on ordinary examples while failing when this assumption is stressed. The likely cause is that training or evaluation rewarded a shortcut. Correct it by adding a targeted counterexample, measuring this failure separately, and tracing the decision back to its source evidence before changing model size.

Notice the shared pattern. Each failure collapses two levels that should remain separate: fluent versus factual, score versus decision, token versus word, correlation versus cause, training distribution versus deployment population, or average quality versus unequal impact. The repair is to restore the missing boundary and measure it directly.

06 Evaluation for a master's-level study

A publishable evaluation begins with a claim table. For every claim, list the dataset slice, metric, baseline, ablation, uncertainty estimate, and known confounder. If the claim is “method A represents long context better,” a single overall accuracy score is insufficient. We need performance by length, a matched-compute baseline, a test that truly requires distant evidence, and an ablation showing the responsible component.

Intrinsic evidence

Does the component optimize or predict what it was designed to model? Useful for fast iteration, but not a substitute for task success.

Extrinsic evidence

Does the representation or model improve a downstream task under a controlled comparison?

Behavioral evidence

Do targeted minimal pairs and adversarial cases show the expected capability rather than a shortcut?

Operational evidence

Are latency, memory, cost, calibration, safety, and subgroup behavior acceptable in the intended environment?

Use a development set for model and threshold choices, then touch the final test set only after the design is fixed. Report variance across seeds when training instability is material. Use grouped or temporal splits when examples share authors, speakers, templates, or evolving events. Deduplicate before splitting. Document preprocessing and evaluate the exact exported pipeline, not an ideal notebook version.

Error analysis should be quantitative enough to change a decision. Sample errors from defined buckets, have more than one reviewer when judgment is subjective, and record disagreement. A confusion matrix, retrieval audit, alignment backtrace, span-boundary table, or per-condition curve is often more actionable than another aggregate benchmark.

07 Study lab: turn the chapter into evidence

These exercises are designed as small research loops. Completing them produces artifacts you can inspect—a derivation, implementation, controlled comparison, and error taxonomy—rather than a vague feeling of familiarity.

01 · Write an explicit preference rubric with examples of ties and uncertainty.

Write the hypothesis before running the exercise. Record the input, expected behavior, metric, and one failure case. Afterward, explain whether the evidence supports the hypothesis and what alternative explanation remains.

02 · Measure inter-annotator agreement and inspect disagreements.

03 · Compare supervised tuning, reward-model scoring, and direct preference objectives conceptually.

04 · Test over-refusal and unsafe compliance as separate axes.

05 · Plot quality versus latency and cost for several test-time-compute budgets.

For an assignment or dissertation notebook, keep a short experiment ledger. Include the question, exact data snapshot, preprocessing hash, model and decoding configuration, random seed, hardware, metric implementation, result, and interpretation. Separate the number you observed from the explanation you infer. That distinction is one of the most valuable habits a master's program can teach.

08 Oral-exam questions

What is the difference between pretraining and instruction tuning?

Pretraining predicts broad text. Instruction tuning concentrates behavior on responding to explicit tasks in a desired format. A strong answer should also state the boundary: which data, task, and assumptions make the claim true, and what observation would cause us to revise it.

Does DPO remove the need for preference data?

No. It simplifies the optimization pipeline, but it still learns from chosen and rejected responses and inherits their coverage and biases. A strong answer should also state the boundary: which data, task, and assumptions make the claim true, and what observation would cause us to revise it.

Does chain-of-thought guarantee reasoning?

No. Longer visible explanations can be wrong or post-hoc. Correctness needs outcome checks, tools, or verifiers when possible. A strong answer should also state the boundary: which data, task, and assumptions make the claim true, and what observation would cause us to revise it.

09 Complete chapter summary

The deeper story is the boundary between a model and a trustworthy language system. Supervised instruction tuning, preference data, reward modeling, RLHF-style optimization, direct preference optimization, alignment limits, reasoning, and test-time compute. Each topic contributes one part of a larger reasoning chain. Representations determine what distinctions are even available. Objectives determine which behavior training rewards. Inference converts scores into a constrained output. Evaluation determines which claim survives contact with held-out evidence. Data connects all four and can quietly invalidate all four through leakage, poor coverage, inconsistent annotation, or historical bias.

Post-training converts general continuation into a more usable assistant policy.
Preference labels are scoped human judgments, not universal truth.
Optimization can exploit weaknesses in the reward signal.
Safety and usefulness need separate, adversarial evaluation.
Test-time compute is valuable when selection and verification are reliable.

Do not leave the chapter with only names of architectures or metrics. Leave with a method: define the task at the grain of the real decision; trace one example from raw input to final output; derive the central computation; preserve uncertainty and provenance; compare against a simple baseline; ablate the claimed contribution; inspect failures by mechanism; and state exactly where the evidence stops. That method will remain useful when today’s architecture is replaced.

copyright and scopeThis is an independent educational companion, not a replacement for the authors' text. It summarizes the chapter's subject in original language and adds new examples, study prompts, and engineering interpretation. For formal definitions, figures, citations, exercises, and the authors' precise treatment, read the official January 6, 2026 source.

Chapter 9 — Post-training: Instruction Tuning, Alignment, and Test-Time Compute.