{"id":"b9350a23e368a57fd1b88e8cdb25d0ea6adc49f2b73b2026ff4552657f66564f","title":"Barak Lenz and co-authors argue in AI21 Labs’ essay for planning actions","argument":"Barak Lenz, Yonatan Belinkov, Subbarao Kambhampati, Kevin Leyton-Brown and Yoav Shoham’s essay was published March 10, 2025. They argue that AI systems should plan sequences of actions rather than rely on language models to predict chains of tokens. An action is an invocation of a tool, such as an LLM prompt, web search or database call. Their proposal puts an external controller around models and tools, choosing and adapting actions according to expected answer quality and costs such as time, compute and money.\n\nThe authors argue that current “Reasoning” Models can be inefficient and may not generalize robustly. They say “thinking tokens” expand the search space, while fixed sampling can waste effort on easier tasks and fall short on harder ones. They also describe limits in user visibility, control and customization, while noting that their account draws on DeepSeek’s R1 and informed guesses about other models.\n\nThey advocate decision-theoretic planning, verification at each step, and interleaving planning with execution. Users, they argue, should be involved throughout and able to correct the system’s understanding of a task. The essay presents results from AI21’s Maestro system, but the retrieved text gives no chart values, so the reported benefits’ size and broader applicability remain unclear.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 27, 2026; originally published March 10, 2025. The page was edited after publication. The retrieved text does not identify which passages, if any, were added or changed after the original publication date. The authors say their account of current LRM approaches is based in part on informed guesses about models other than DeepSeek’s R1. The retrieved text refers to Maestro charts but contains no chart values, limiting assessment of the reported results.","speaker":{"name":"AI21 Labs","roleAtSource":"Publisher"},"source":{"title":"Planning Actions, Not Predicting Tokens","publisher":"AI21 Labs","format":"Article","url":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","publishedAt":"2025-03-10T18:27:35Z","dateBasis":"Date from the publisher’s own records, through its API."},"evidence":[{"text":"AI21 Labs’ co-authors argue that systems should search sequences of tool-use actions rather than token sequences.","excerpt":"The space that enables systematic search for solutions is not the space of token sequences, but the space of action sequences.","url":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","relationship":"Source-stated"},{"text":"The authors say planning should account for distributions over answer quality and costs, including time, compute and money.","excerpt":"taking into account distributions over both answer quality and cost (in terms of time, compute and money).","url":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","relationship":"Source-stated"},{"text":"The authors acknowledge that their discussion of existing LRMs is based on DeepSeek’s R1 and informed guesses about other models.","excerpt":"this is based on the one model – DeepSeek’s R1 – that was described in a fair amount of detail, and informed guesses about the rest.","url":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","relationship":"Source-stated"},{"text":"The authors report that they present results from Maestro, but the retrieved text supplies no chart values to assess those results.","excerpt":"Our goal was to describe the ideas rather than the specific offering of AI21, but to get a concrete sense for the benefits, here are some results from Maestro, our planning and orchestration system.","url":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well would this planning approach perform on tasks and tools beyond the examples discussed in the essay?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Barak Lenz and co-authors argue in AI21 Labs’ essay for planning actions","summary":"Barak Lenz, Yonatan Belinkov, Subbarao Kambhampati, Kevin Leyton-Brown and Yoav Shoham’s essay was published March 10, 2025. They argue that AI systems should plan sequences of actions rather than rely on language models to predict chains of tokens. An action is an invocation of a tool, such as an LLM prompt, web search or database call. Their proposal puts an external controller around models and tools, choosing and adapting actions according to expected answer quality and costs such as time, compute and money.\n\nThe authors argue that current “Reasoning” Models can be inefficient and may not generalize robustly. They say “thinking tokens” expand the search space, while fixed sampling can waste effort on easier tasks and fall short on harder ones. They also describe limits in user visibility, control and customization, while noting that their account draws on DeepSeek’s R1 and informed guesses about other models.\n\nThey advocate decision-theoretic planning, verification at each step, and interleaving planning with execution. Users, they argue, should be involved throughout and able to correct the system’s understanding of a task. The essay presents results from AI21’s Maestro system, but the retrieved text gives no chart values, so the reported benefits’ size and broader applicability remain unclear.","question":"How well would this planning approach perform on tasks and tools beyond the examples discussed in the essay?","claims":[{"text":"AI21 Labs’ co-authors argue that systems should search sequences of tool-use actions rather than token sequences.","excerpt":"The space that enables systematic search for solutions is not the space of token sequences, but the space of action sequences.","support":"source-stated"},{"text":"The authors say planning should account for distributions over answer quality and costs, including time, compute and money.","excerpt":"taking into account distributions over both answer quality and cost (in terms of time, compute and money).","support":"source-stated"},{"text":"The authors acknowledge that their discussion of existing LRMs is based on DeepSeek’s R1 and informed guesses about other models.","excerpt":"this is based on the one model – DeepSeek’s R1 – that was described in a fair amount of detail, and informed guesses about the rest.","support":"source-stated"},{"text":"The authors report that they present results from Maestro, but the retrieved text supplies no chart values to assess those results.","excerpt":"Our goal was to describe the ideas rather than the specific offering of AI21, but to get a concrete sense for the benefits, here are some results from Maestro, our planning and orchestration system.","support":"source-stated"}],"limitations":["Text as retrieved on October 27, 2026; originally published March 10, 2025.","The page was edited after publication. The retrieved text does not identify which passages, if any, were added or changed after the original publication date.","The authors say their account of current LRM approaches is based in part on informed guesses about models other than DeepSeek’s R1.","The retrieved text refers to Maestro charts but contains no chart values, limiting assessment of the reported results."],"mediaId":null,"correctionReason":null,"id":"b9350a23e368a57fd1b88e8cdb25d0ea6adc49f2b73b2026ff4552657f66564f","publicationId":"a1501448-4abd-8ad5-8470-31d934f33f13","version":1,"sourceUrl":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","sourceTitle":"Planning Actions, Not Predicting Tokens","publisher":"AI21 Labs","publishedAt":"2025-03-10T18:27:35Z","originalPublished":{"value":"2025-03-10T18:27:35Z","precision":"instant","basis":"backfill:api_published"},"firstObservedAt":"2026-10-27T15:22:13.212Z","createdAt":"2026-11-14T23:53:23.480Z","historical":true,"entityIds":["ai21"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"99a58fe24a267ea0171f70c135adab3a7681d2433607bd07e34038d5c6b36c2c","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-27T15:22:08.831Z"},"provenance":{"textHash":"b9cd202c9626fd37d69dcd92c1babc36790a52575aab3f274f07c0004cd2da2d","observationId":"5acc3f474b7fadbdd481f092f2b9e8bf63bccc861601d3a58c75eda560378aa5","retrievedAt":"2026-10-27T15:23:38.689Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"43c5dec7-c2e7-48e3-a09f-03e9ac7bd0a5"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"b429e667-a7f1-7357-260a-c5bf8703cdb3"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-14T23:48:44.342Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/barak-lenz-and-co-authors-argue-in-ai21-labs-essay-for-planning-b9350a23","formats":{"html":"/readings/barak-lenz-and-co-authors-argue-in-ai21-labs-essay-for-planning-b9350a23","markdown":"/readings/barak-lenz-and-co-authors-argue-in-ai21-labs-essay-for-planning-b9350a23.md","json":"/readings/barak-lenz-and-co-authors-argue-in-ai21-labs-essay-for-planning-b9350a23.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"conditions":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-llm","name":"Large language model","explanation":"A language model with many learned numerical settings, used to process and generate language.","referenceUrl":"https://developers.google.com/machine-learning/glossary#large-language-model"}],"entities":[{"id":"ai21","name":"AI21","kind":"organization","context":"AI21 publishes technical research and product announcements about its language models and AI systems.","referenceUrls":["https://www.ai21.com/blog/"],"path":"/entities/ai21","json":"/entities/ai21.json","markdown":"/entities/ai21.md"},{"id":"time","name":"TIME","kind":"organization","context":"TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.","referenceUrls":["https://time.com/"],"path":"/entities/time","json":"/entities/time.json","markdown":"/entities/time.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"b9350a23e368a57fd1b88e8cdb25d0ea6adc49f2b73b2026ff4552657f66564f","revision":"a1501448-4abd-8ad5-8470-31d934f33f13","revisionKind":"publication-id","title":"Barak Lenz and co-authors argue in AI21 Labs’ essay for planning actions","url":"https://escapemeridian.com/readings/barak-lenz-and-co-authors-argue-in-ai21-labs-essay-for-planning-b9350a23"},"source":{"type":"source","id":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","revision":"b9cd202c9626fd37d69dcd92c1babc36790a52575aab3f274f07c0004cd2da2d","revisionKind":"sha256-source-text","title":"Planning Actions, Not Predicting Tokens","url":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"b9350a23e368a57fd1b88e8cdb25d0ea6adc49f2b73b2026ff4552657f66564f","title":"Barak Lenz and co-authors argue in AI21 Labs’ essay for planning actions","argument":"Barak Lenz, Yonatan Belinkov, Subbarao Kambhampati, Kevin Leyton-Brown and Yoav Shoham’s essay was published March 10, 2025. They argue that AI systems should plan sequences of actions rather than rely on language models to predict chains of tokens. An action is an invocation of a tool, such as an LLM prompt, web search or database call. Their proposal puts an external controller around models and tools, choosing and adapting actions according to expected answer quality and costs such as time, compute and money.\n\nThe authors argue that current “Reasoning” Models can be inefficient and may not generalize robustly. They say “thinking tokens” expand the search space, while fixed sampling can waste effort on easier tasks and fall short on harder ones. They also describe limits in user visibility, control and customization, while noting that their account draws on DeepSeek’s R1 and informed guesses about other models.\n\nThey advocate decision-theoretic planning, verification at each step, and interleaving planning with execution. Users, they argue, should be involved throughout and able to correct the system’s understanding of a task. The essay presents results from AI21’s Maestro system, but the retrieved text gives no chart values, so the reported benefits’ size and broader applicability remain unclear.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 27, 2026; originally published March 10, 2025. The page was edited after publication. The retrieved text does not identify which passages, if any, were added or changed after the original publication date. The authors say their account of current LRM approaches is based in part on informed guesses about models other than DeepSeek’s R1. The retrieved text refers to Maestro charts but contains no chart values, limiting assessment of the reported results.","speaker":{"name":"AI21 Labs","roleAtSource":"Publisher"},"source":{"title":"Planning Actions, Not Predicting Tokens","publisher":"AI21 Labs","format":"Article","url":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","publishedAt":"2025-03-10T18:27:35Z","dateBasis":"Date from the publisher’s own records, through its API."},"evidence":[{"text":"AI21 Labs’ co-authors argue that systems should search sequences of tool-use actions rather than token sequences.","excerpt":"The space that enables systematic search for solutions is not the space of token sequences, but the space of action sequences.","url":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","relationship":"Source-stated"},{"text":"The authors say planning should account for distributions over answer quality and costs, including time, compute and money.","excerpt":"taking into account distributions over both answer quality and cost (in terms of time, compute and money).","url":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","relationship":"Source-stated"},{"text":"The authors acknowledge that their discussion of existing LRMs is based on DeepSeek’s R1 and informed guesses about other models.","excerpt":"this is based on the one model – DeepSeek’s R1 – that was described in a fair amount of detail, and informed guesses about the rest.","url":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","relationship":"Source-stated"},{"text":"The authors report that they present results from Maestro, but the retrieved text supplies no chart values to assess those results.","excerpt":"Our goal was to describe the ideas rather than the specific offering of AI21, but to get a concrete sense for the benefits, here are some results from Maestro, our planning and orchestration system.","url":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well would this planning approach perform on tasks and tools beyond the examples discussed in the essay?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Barak Lenz and co-authors argue in AI21 Labs’ essay for planning actions","summary":"Barak Lenz, Yonatan Belinkov, Subbarao Kambhampati, Kevin Leyton-Brown and Yoav Shoham’s essay was published March 10, 2025. They argue that AI systems should plan sequences of actions rather than rely on language models to predict chains of tokens. An action is an invocation of a tool, such as an LLM prompt, web search or database call. Their proposal puts an external controller around models and tools, choosing and adapting actions according to expected answer quality and costs such as time, compute and money.\n\nThe authors argue that current “Reasoning” Models can be inefficient and may not generalize robustly. They say “thinking tokens” expand the search space, while fixed sampling can waste effort on easier tasks and fall short on harder ones. They also describe limits in user visibility, control and customization, while noting that their account draws on DeepSeek’s R1 and informed guesses about other models.\n\nThey advocate decision-theoretic planning, verification at each step, and interleaving planning with execution. Users, they argue, should be involved throughout and able to correct the system’s understanding of a task. The essay presents results from AI21’s Maestro system, but the retrieved text gives no chart values, so the reported benefits’ size and broader applicability remain unclear.","question":"How well would this planning approach perform on tasks and tools beyond the examples discussed in the essay?","claims":[{"text":"AI21 Labs’ co-authors argue that systems should search sequences of tool-use actions rather than token sequences.","excerpt":"The space that enables systematic search for solutions is not the space of token sequences, but the space of action sequences.","support":"source-stated"},{"text":"The authors say planning should account for distributions over answer quality and costs, including time, compute and money.","excerpt":"taking into account distributions over both answer quality and cost (in terms of time, compute and money).","support":"source-stated"},{"text":"The authors acknowledge that their discussion of existing LRMs is based on DeepSeek’s R1 and informed guesses about other models.","excerpt":"this is based on the one model – DeepSeek’s R1 – that was described in a fair amount of detail, and informed guesses about the rest.","support":"source-stated"},{"text":"The authors report that they present results from Maestro, but the retrieved text supplies no chart values to assess those results.","excerpt":"Our goal was to describe the ideas rather than the specific offering of AI21, but to get a concrete sense for the benefits, here are some results from Maestro, our planning and orchestration system.","support":"source-stated"}],"limitations":["Text as retrieved on October 27, 2026; originally published March 10, 2025.","The page was edited after publication. The retrieved text does not identify which passages, if any, were added or changed after the original publication date.","The authors say their account of current LRM approaches is based in part on informed guesses about models other than DeepSeek’s R1.","The retrieved text refers to Maestro charts but contains no chart values, limiting assessment of the reported results."],"mediaId":null,"correctionReason":null,"id":"b9350a23e368a57fd1b88e8cdb25d0ea6adc49f2b73b2026ff4552657f66564f","publicationId":"a1501448-4abd-8ad5-8470-31d934f33f13","version":1,"sourceUrl":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","sourceTitle":"Planning Actions, Not Predicting Tokens","publisher":"AI21 Labs","publishedAt":"2025-03-10T18:27:35Z","originalPublished":{"value":"2025-03-10T18:27:35Z","precision":"instant","basis":"backfill:api_published"},"firstObservedAt":"2026-10-27T15:22:13.212Z","createdAt":"2026-11-14T23:53:23.480Z","historical":true,"entityIds":["ai21"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.ai21.com/blog/ai-reasoning-planning-vs-predicting/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"99a58fe24a267ea0171f70c135adab3a7681d2433607bd07e34038d5c6b36c2c","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-27T15:22:08.831Z"},"provenance":{"textHash":"b9cd202c9626fd37d69dcd92c1babc36790a52575aab3f274f07c0004cd2da2d","observationId":"5acc3f474b7fadbdd481f092f2b9e8bf63bccc861601d3a58c75eda560378aa5","retrievedAt":"2026-10-27T15:23:38.689Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"43c5dec7-c2e7-48e3-a09f-03e9ac7bd0a5"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"b429e667-a7f1-7357-260a-c5bf8703cdb3"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-14T23:48:44.342Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"a1501448-4abd-8ad5-8470-31d934f33f13","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading correctly credits the five named co-authors and AI21 Labs as publisher. It also keeps their criticisms of LRMs and their Maestro results attributed to the authors rather than presenting them as established fact.","excerpt":"Co-authored by Barak Lenz, Yonatan Belinkov, Subbarao Kambhampati, Kevin Leyton-Brown, Yoav Shoham."},{"dimension":"summary","status":"uncertain","explanation":"The reading is mostly faithful, but it has two fixable gaps. First, it softens the authors' force: they assert LRMs have an 'inability to generalize in a robust way' and call them free associators rather than reasoners, but the reading says they 'may not generalize robustly'. Repair: state the claim as the authors make it. Second, 'tokens' and 'thinking tokens' are central and are not defined in the reading or its definitions list, and 'LRMs' in the claims is never expanded. Repair: add a short plain-language gloss or tooltip for each.","excerpt":"A separate, and perhaps deeper, algorithmic limitation of LRMs is their inability to generalize in a robust way."},{"dimension":"date","status":"pass","explanation":"The reading gives March 10, 2025, which is the UTC calendar day of the 2025-03-10T18:27:35Z instant. It uses no hindsight framing and does not rely on any later passages; the 'closed preview' note is not used. The evidence is a publisher page, not an archive copy. The reading's limitations already disclose the post-publication edit.","excerpt":null},{"dimension":"support","status":"pass","explanation":"Each claim matches its exact excerpt. The summary's points on fixed K sampling, thinking tokens enlarging the search space, step-level verification, interleaving planning with execution, and user involvement are all source-stated. The reading notes that the retrieved text contains no chart values rather than inflating Maestro's results. The page was edited later, but the cited passages read as integral to the original argument.","excerpt":"K is constant across all inputs, which is inefficient: for easy cases it’s an overkill, for hard cases it’s not enough"},{"dimension":"media","status":"pass","explanation":"No media is selected and no verified candidates were supplied, so nothing is misidentified.","excerpt":null}],"createdAt":"2026-11-14T23:48:44.342Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}