1The question is what changed

RSI: Recursive self-improvement: improvements in the system help produce new improvements. The acronym comes from the English recursive self-improvement.

An AI can rewrite a message without changing how it works. It can also propose a new work rule, saved for future use. These are different changes. To talk about recursion, look for evidence that the new version helps improve the next versions.

In Marina’s fictional stationery shop, shortening a reply changes the text. Saving a comparison procedure changes future work.

2Local improvements have value

You can take advantage of improvement cycles without building an autonomous AI. In this course, the practices are supervised experiments. They teach you to propose, compare, and decide. They don’t show an intelligence explosion or automatically change the model used in the chat.

Ícaro, a fictional educator, compares instructions to create exercises. He saves the best instruction, but he doesn’t claim he trained a new AI.

3The cycle needs evidence

A version that praises itself does not prove improvement. Record the previous version, the test, the result, and who decided. Separate capability in a task from the ability to do research. More speed in one step may not change the performance of the whole.

Marina keeps the old and new answer. She only changes the guidance after testing questions she didn’t use in the review.

1Context is the material from this attempt

parameters: Values internally adjusted during training. They are not the messages you write in a conversation.

The model answers using available instructions and information. A correction in the conversation can change the next response. That does not mean the model’s parameters were updated. When you start another conversation, part of that material may no longer be available.

Ícaro adds a short text about recycling before asking questions. The answer improves because it received the necessary reference.

2Memory needs to be consulted and checked

Saving a rule helps only when the system retrieves it at the right time. The rule can grow outdated or contradict another. Name the record, add a date, and specify who is responsible. Before using important information, confirm whether it still holds.

Marina keeps customer service guidance in a document. When the store’s hours change, she updates the guidance and checks old answers.

3Training is another intervention

Training adjusts parameters with examples or evaluation signals. Changing tools or instructions can improve the system without that adjustment. Start with the simplest layer that solves the problem. Don’t use training to compensate for a missing reference.

Ícaro wants the questions to cite the supporting excerpt. He first changes the instruction and measures the result; he does not need to start by training a model.

1Start with the exact statement

evidence: A record that lets you verify a claim. It may include an experiment, its measurements, and its conditions.

Separate a verifiable sentence from the opinion that comes with it. Look for the original study or report, its date, and its authors. A video can present good ideas, but its narration does not replace the original document. Also record whether there was external evaluation.

Marina reads “AI does all the research”. She looks for which tasks were done and which decisions stayed human.

2Activity is not the final result

More hours of execution or more experiments do not directly measure useful discoveries. In September 2026, OpenAI reported greater internal use of agents. It also said that people still choose priorities and decide which results to pursue.

Ícaro can generate a hundred exercises in an afternoon. That does not tell you how many are correct or help his students.

3Write the limit next to the result

An experiment can have few cases or very specific tasks. A future scenario is a tool for discussing possibilities. It does not establish a guaranteed timeline. Use three labels in your record: observed, interpretation, and prediction. Leave what you couldn’t verify explicitly open.

Marina agrees to test a new routine. She does not change the whole company because of a predicted deadline.

1The reference comes before the proposal

LOOP-R: Educational adaptation of the LOOP-R method: execute, measure, critique, propose, test, validate, promote, and repeat.

Execute the task with the current guidance and save the result. Define a simple measure tied to the goal. Only then ask for a proposal. Without an initial reference, a convincing answer can seem better just because it is new. Keep the same cases in the comparison.

Marina records whether each answer provides a time, avoids promises without a basis, and indicates the next action.

2Test one change at a time

Criticizing means finding an observable flaw. Proposing means suggesting an adjustment that can fix it. Testing means running again under comparable conditions. If you change instruction, model, and examples all together, it becomes hard to point to a single cause for the result.

Ícaro adds only “cite the supporting excerpt.” He keeps the text, the number of questions, and the previous criteria unchanged.

3Promoting creates a reference for future uses

Validation checks the candidate on held-out cases. Approval authorizes its adoption within written limits. Promotion records that decision and makes the approved version current. In equivalent future uses, the system applies the approved rule without asking again. A new candidate still needs assessment and a decision.

Marina approves the schedule rule for drafts. The team reuses that rule as long as the source and authorization remain valid.

1Define the inputs and outputs for each role

evaluator: Person or mechanism that applies quality criteria to a result. It needs evidence, not just opinion.

A proposer receives the task, the observed errors, and the limits. They return a candidate and the justification. The evaluator receives cases and criteria, and returns checkable results. The approver reviews quality, cost, and undesired effects before deciding.

Ícaro asks the AI for a new instruction, checks the questions against the text, and decides whether to use the instruction.

2Two conversations don’t guarantee independence

Separating screens helps you organize the work, but models can repeat the same mistakes. Use verifiable criteria and examples the proposal didn’t receive. For factual statements, check the reference. For subjective results, record disagreements instead of hiding them in a score.

Marina asks the other conversation for criticism. Even so, she compares the time with the store’s document.

3Approval applies to a written scope

Whoever approves sets the rule, the permitted uses, and the limits. A recorded approval keeps working for equivalent uses. Don’t ask again just because another run started. Changing the rule or expanding your actions requires a new decision. Approving content does not automatically authorize sending messages.

Ícaro approves a rule for preparing drafts. The system reuses the rule, but it does not change grades or contact families.

1Record what was truly approved

approved knowledge: Rule or information accepted for defined uses, with source, version, responsible person, and validity. Approval doesn’t make it infallible.

Separate the proposal, the test result, and the approved rule. In LOOP-R, the promotion links the decision to the current version. Record the exact content, the source, the tests, and who authorized it. Write which tasks the approval applies to. Without that, the system can’t tell approved knowledge apart from an attempt.

Marina approves rule H1: use the current policy’s time in drafts. Holidays still have no information.

2Reuse without reopening the same decision

Before using it, the system retrieves the current rule and checks the scope. This check is not a request for approval. If task, source, and validity stay compatible, it continues without asking again. An approved preference can guide dozens of answers; each answer is still subject to the agreed criteria.

Ícaro approves citing the reference excerpt in the questions. The system applies this rule to new drafts without asking for the same authorization.

3Approve with future uses in mind

An approval can power dozens of future decisions. Check content, evidence, coverage, and the cost of being wrong before approving. If the source changes, a contradiction appears, or the validity ends, pause the affected use and review it. Revoking disables the rule, but preserves the history. Human approval can also be wrong.

Policy P2 changes the time to 10h–17h. Marina suspends H1 until she reviews the reference; the system doesn’t insist on the old time.

1Propose and verify work together

bottleneck: Stage that limits overall performance. Improving another stage can bring little total gain.

AlphaEvolve combines language models, automatic evaluations, and search among candidate programs. The published results include computational optimizations. The problem needs to allow a reliable evaluation. A sleek proposal that produces wrong answers shouldn’t be chosen just because it’s fast.

Marina makes an analogy with the counter: answering faster doesn’t help when the information is wrong.

2The percentage belongs to a measurement

DeepMind reported a 23% speedup in an operation used by Gemini. The reported effect on training time was 1%. These values refer to different scales. The case shows why you should ask where the gain happened.

Icaro takes less time to format questions. You still need to read the text and check the answers.

3Transfer the logic, not the promise

In your work, choose a task whose quality can be checked. Define the correct result before measuring speed. Compare total effort, including review. The usefulness of this logic does not depend on having access to the lab’s search tool.

Marina measures the time between receiving the question and approving the answer. She does not measure only the time the chat takes to write.

1 The altered object is the agent

evolutionary search: Exploration of variants, with evaluation and selection across rounds. It doesn’t require living organisms.

The DGM uses models to propose changes to the code that organizes the agent. This includes tools and resolution procedures. It does not mean retraining all the language model’s parameters again. Identifying the object helps you avoid confusing a system improvement with a model swap.

In Ícaro’s analogy, changing the checking routine changes the work. It does not automatically change the teacher’s training.

2 Saved variants open new paths

The study keeps a growing collection of agents. A new attempt can start from different variants. A mid-level candidate might contain an idea that helps another round. Preserving diversity doesn’t mean using worse candidates in real life.

Marina saves a short answer and a detailed one between attempts. None becomes the current guidance without passing the criteria.

3 Test result isn’t a universal guarantee

The authors report gains in programming evaluations. The reach depends on the tasks and conditions studied. They also document cases where the evaluation was fooled. That’s why a single score doesn’t end the investigation. Look at the result and the path used to get it.

Ícaro checks whether the question has an answer in the text. A “pass” mark generated by the candidate itself isn’t enough.

1Automating scientific steps doesn’t solve all of science

isolated environment: Execution space with defined limits for a task. Its protection depends on the implementation and permissions.

Sakana reported a piece of work by the AI Scientist accepted in a workshop connected to ICLR 2025. The workshop agreed to participate in the evaluation experiment. This is not the same as acceptance at the main conference. An accepted article also doesn’t prove that every conclusion is correct.

Ícaro distinguishes a draft of an activity from its validation in the classroom. Each step adds different evidence.

2Experiments need a place to happen

DSec is an infrastructure described by DeepSeek for running agent tasks at scale. Preparing environments, controlling resources, and recording actions are concrete problems. Efficient infrastructure enables more tests. It doesn’t, by itself, prove that a system chooses good questions or improves them recursively.

Marina reserves fictional copies to test responses. Preparing the workstation makes the work easier, but it doesn’t determine which response is good.

3Notice who decides the direction

Fast research reports don’t remove the need to choose goals. Ask who defines the next experiment and who can stop it. Counting attempts is easy. Measuring scientific contribution means assessing what each attempt actually added.

Ícaro can generate many exercise variations. He keeps deciding which ones meet the lesson’s goal.

1Choose cases that represent the work

held-out cases: Examples separated before the changes and used afterward to evaluate the candidate. They shouldn’t guide every review.

Include common situations, missing data, and requests out of scope. Define the expected response or the criteria before running. A small sample teaches the method, but it doesn’t prove overall performance. Avoid choosing only easy cases after you’ve seen the results.

Marina includes one question with no date and another outside the provided time window. This tests whether the response asks for clarification.

2Keep part of the cases before adjusting

Use one group to improve guidance. Set another group aside to check the chosen version. After you consult the held-out results repeatedly, they no longer provide an independent check on unseen cases. For future decisions, prepare other cases and record this replacement.

Ícaro adjusts the instruction using two texts. A third one stays saved until the final comparison.

3Record more than an average score

Mark criteria per case, in addition to time and review effort. An average can hide an important failure. Repeat attempts when answers vary. Compare versions under similar conditions. Don’t turn a small test into a promise of accuracy for any task.

One of Marina’s responses was short, but it invented a condition for exchanging a purchase. This error prevents approval even if the response was fast.

1An incomplete goal creates shortcuts

reward manipulation: Get a high score by exploiting flaws in the assessment, without fulfilling the task’s intent.

If the goal is only to answer quickly, leaving out information can seem like a win. If it’s to increase quantity, duplicating results can seem like progress. Also write what must stay correct. Combine the main measure with criteria that prevent these losses.

Marina won’t accept empty messages to reduce customer service time. The response must resolve the question or ask for the missing information.

2Statements are not execution

A tool can claim it checked something without producing evidence. Keep the observable result of the check. Distinguish a teaching simulation from real execution. When there isn’t enough proof, write “not verified” instead of turning trust into approval.

Ícaro asks for the excerpt that supports each answer. “Everything checked” doesn’t replace finding the sentences.

3Protect the criteria from convenient changes

A candidate system should not make its own test easier. Preserve the cases, the answer key, and the original results. If a rule is wrong, revise it in a separate process and reevaluate all versions. Don’t change criteria just to save your preferred candidate.

Marina finds an ambiguous criterion. She corrects it and applies it again to both versions, without favoring the new one.

1Start with what the examples represent

synthetic data: Examples produced artificially, including by AI. They can serve for tests and training, with proper verification.

Choose categories needed for the task. Include common situations and relevant exceptions. Mark which examples were generated and who verified them. A collection of nearly identical variations increases volume without greatly expanding coverage.

Ícaro notices that ten generated questions ask for the same information. He swaps part of them for questions that assess other ideas from the text.

2Don’t feed the cycle only with your own outputs

Studies show degradation under certain conditions of recursive training with generated data. This doesn’t mean that all synthetic data causes collapse. The composition, selection, and preservation of diversity matter. Don’t replace verified references with answers generated without checking.

Marina uses the schedule document as a reference. She doesn’t treat an old chat response as the official source.

3AI evaluation is limited support

An AI can compare results and locate problems. But it may prefer long text or agree with plausible errors. Have a person check a sample using objective criteria. When the judgment is uncertain, keep the disagreement instead of fabricating a conclusion.

Icaro uses automatic criticism to locate suspicious questions. He checks the text before accepting the evaluation.

1Provide a small, complete reference

fidelity: Correspondence between the response and the available reference. A faithful response doesn’t add facts without support.

Use a fictional policy to learn the method. Define which information is available. Ask the response to acknowledge gaps. A cordial tone instruction doesn’t replace facts. The first test should reveal whether the system distinguishes known information from unknown information.

Marina tests an imaginary stationery store. The reference includes opening hours and pickup, but it doesn’t mention the exchange period.

2Improve only the response guidance

Compare a basic instruction with another that requires indicating missing data. Keep the model and the reference. Use separate conversations or clearly identify each attempt. If the tool isn’t available, manually compare the provided drafts.

Ícaro applies the same logic to school announcements. A lack of information doesn’t turn into imagined authorization.

3Validate with a different question

After adjusting the instruction, test a reserved question. Look for made-up facts and confusing actions. Also note whether the answer resolves what it can resolve. Saying “I don’t know” to everything avoids inventions, but it can make the service useless.

Marina asks when she can pick up an order. The answer should use the available condition: after the notice.

1Tell what the student should demonstrate

quality criterion: Observable rule to decide whether the result meets the learning objective. It must be defined before comparing versions.

Choose a small skill, like locating explicitly stated information. Don’t mix this goal with external research without telling. Provide the text and the number of questions. The assessment must match the chosen skill, not just the look of a school activity.

Ícaro wants the class to locate recycling conditions. He uses a short text and does not require outside knowledge.

2Ask for evidence with the answer

Ask for the question, the answer, and the supporting excerpt. Check whether the excerpt truly supports the answer. Copying any sentence doesn’t solve it. Require the tool to discard questions without support in the material. You can apply this procedure by hand, without depending on a chat.

Marina applies the same logic in a team training. Each question must point back to the guidance that was provided.

3Check variety and appropriateness

Three almost identical questions don’t assess three different ideas. Check repetition, clarity, and difficulty. The tool may suggest a revision, but it doesn’t automatically know the class. For a next round, use another reserved text and keep the same criteria.

Ícaro rejects two questions that ask about the same day. He adds a question about the condition of the paper.

1Separate extraction from suggestion

traceability: Possibility of linking each statement to the material that supports it and to the record of how it was produced.

Extraction recovers information that is already present. Suggesting adds a possibility. Both actions can be useful, as long as they’re identified. Ask for a factual list first. If you want suggestions, put them in a separate part, without presenting them as decisions already made.

Marina wants to turn meeting notes into actions. She doesn’t want the AI to choose deadlines to fill in empty spaces.

2A blank (missing piece of info) is a valid result

Use “not provided” when the reference doesn’t contain the data. Don’t confuse that label with a lack of capability. It shows the limit of the material. A list with explicit blanks can be more useful than a complete, invented summary.

Ícaro prefers to receive “pending deadline”. That way, they know which question to ask before planning the activity.

3Compare item by item

Evaluate person, action, and deadline separately. Count omissions and additions without support. Then check clarity. When you review the guidance, change only one rule. Keep the original note available so another person can verify the list.

Marina checks each line against the note. The review finds the date copied from another task incorrectly.

1A chat and a worksheet are enough to get started

autonomy: How much a system can decide and carry out without human intervention. It needs clear scope and limits.

In your first attempts, copy the material, compare answers, and record results. This makes the method visible. You learn which cases fail before you expand execution. You don’t need to buy a special tool to do the exercises in this course.

Marina compares two approaches on a sheet. She only considers automation after she understands which failures she needs to detect.

2The next level needs requirements

An automatic routine needs to receive data, run proposals, evaluate, and record. Someone maintains the limits and handles failures. Before you hire or build, write down these needs. The provider must show how the result can be checked and how to go back to the previous version.

Ícaro specifies that the routine only prepares questions. The educational approval stays with him.

3Frontier research isn’t a button

Modifying agents and running training involves specialization, extra resources, and additional risks. Don’t treat lab examples as a ready-made recipe for any company. Choose the smallest level that meets the need. Local gains can be useful without broad autonomy.

Marina doesn’t need to build a lab to reduce service mistakes. First, validate the guidance used in the drafts.

1Calculate the cost of the usable result

stop condition: Event that ends the test, like reaching the limit of attempts, going over cost, or detecting a critical failure.

Add time for preparation, execution, and review. If there’s a charge based on usage, record the observed value. Don’t confuse the advertised price with the total cost of the work. Compare recurring savings with the effort to prepare and maintain the change.

Marina estimates how much time she saves per response and how much she spent preparing the test. A rare improvement may not cover the preparation.

2Write simple, verifiable limits

For a first test, choose a few cases and up to three candidates. Set a work window. Define a financial cap if the tool charges. Record educational examples and values that were truly measured separately. Don’t move the exercise budget to any project.

Ícaro reserves a session to compare instructions. If there isn’t time to check, he doesn’t approve the candidate out of haste.

3The reversal is part of the plan

Keep the previous guidance and indicate who can restore it. Stop when you come across invented relevant information, an unauthorized action, or a cost above the ceiling. Investigate the failure before repeating. If the system changes, reassess the pilot conditions.

Marina keeps the current guidance. A candidate who creates improper commitments is rejected before reaching the team.

1Choose a clearly defined task

validation: Candidate review with defined criteria and held-out cases, before deciding whether to use it.

Use support services, reading questions, or action extraction. If you didn’t do the previous practices, copy one of the fictional materials from this module. Write the initial reference and one change. You can run the trial with an AI chat or compare drafts manually.

Ícaro chooses questions that can be answered by the text. His only change is requiring a supporting passage.

2Run and record the comparison

Apply both guidelines to the same cases. Keep the responses, the criteria, and the observed times. Then use the held-out cases without continuing to adjust the candidate. If there’s any variation, record it. A small difference in a short trial can be inconclusive.

Marina tests common questions and gaps. She keeps a worse answer alongside the best ones, so she doesn’t select only wins.

3Also decide what can be reused

Conclude: approve for limited use, reject, or investigate more. If you approve, record the accepted knowledge, the version, and the future authorized uses. The same valid rule doesn’t require approval every time you repeat it. New changes go through review. State what the test didn’t prove and how to suspend or revoke the decision.

Icaro approves a rule for drafts with pedagogical review. This rule does not need new approval for every activity.