Subbarao Kambhampati: planning, explanation and the discipline of verification

Biographies · Planning, human–AI collaboration & verification

A plan must survive contact with the conditions it depends on. A journey needs workable connections; a robot must know what an action will change. Subbarao Kambhampati has spent decades studying how machines can construct such plans, explain them to people and establish whether they will work. His recent criticism of language models grows from that longer inquiry into the relationship between knowledge, action and justified confidence.

Learning to revise a plan

Kambhampati studied electrical engineering and electronics at the Indian Institute of Technology Madras, graduating in 1983. At the University of Maryland, he completed a master’s degree in computer science in 1985 and a doctorate in 1989. His doctoral research with James Hendler concerned the flexible reuse and modification of hierarchical plans: how an existing course of action might be adapted when circumstances change. The problem already required attention to both what could be preserved and what had to be reconsidered.

After working as a research associate at Stanford, he joined Arizona State University in 1991 and became a full professor in 2000. His academic record traces a career in which planning gradually connected with questions about people’s expectations and knowledge. At ASU, where he leads the Yochan research group, the collaboration between humans and intelligent systems became a central research concern.

Time, resources and the cost of acting

With Minh B. Do, Kambhampati developed Sapa, described in the Journal of Artificial Intelligence Research in 2003. The planner handled actions that take time, consume resources and must meet deadlines. Its search was guided by estimates sensitive to both cost and the time needed to complete a plan. That combination matters whenever a cheaper route takes longer, or several actions compete for a limited resource.

Sapa illustrates the constructive foundation of his later arguments. Planning requires a representation of actions and their consequences, together with a method for searching possible futures. The system’s effectiveness depended on its models and approximations. Its contribution was a concrete approach to those difficulties, within the problems it was designed to represent.

Explaining a decision to another mind

A technically valid plan can still puzzle the person expected to accept it. In a 2017 IJCAI paper, Tathagata Chakraborti, Sarath Sreedharan, Yu Zhang and Kambhampati treated explanation as a problem of reconciling models. The person and the machine may have different information about the same situation. An explanation should address the difference that makes a decision appear unreasonable.

Consider a robot taking a longer route because a passage is blocked. Someone unaware of the obstruction needs that missing fact to understand the detour. The example captures the idea: explain the relevant difference in knowledge. In the paper’s formal account, changes to the human’s model make the proposed plan optimal under that revised model. Applying this approach requires representing what the person knows, which introduces assumptions of its own. Explanation is consequently a matter of communication as well as computation.

Putting language-model plans to the test

The arrival of conversational AI brought a new question to this research tradition: could a fluent model reliably organise actions? In 2023, Karthik Valmeekam, Matthew Marquez, Alberto Olmo, Sarath Sreedharan and Kambhampati introduced PlanBench. Drawing on established planning domains, the benchmark makes tasks about actions and changing states sufficiently explicit for answers to be checked.

A companion investigation, by Valmeekam, Marquez, Sreedharan and Kambhampati, found limited success when the tested models generated executable plans autonomously. It also found more promising uses for their suggestions in guiding conventional planners, with external verification helping improve generated plans. These were results for particular models and experimental conditions in 2023. They establish a dated comparison against which subsequent progress can be measured.

Giving language models a verifiable role

With Lin Guan, Valmeekam and Sreedharan, Kambhampati also investigated using language models to construct explicit models of actions. In their 2023 work, a model helped express a domain in a formal planning language and incorporate corrective feedback from people and validators. A separate planner then used the corrected representation. Here, a “world model” means an explicit account of actions and conditions; it differs from learned visual predictors or the latent dynamics used by systems such as Dreamer.

In an ICML 2024 position paper, Kambhampati and seven collaborators set out LLM-Modulo. A language model proposes candidates; external components check them and return feedback. The model can also help acquire knowledge and clarify requirements. The proposal gives language generation a useful place in a broader system while locating formal correctness in the checking procedures.

Those guarantees have boundaries. A verifier checks the conditions represented in its model; an omitted constraint can still matter in practice. A procedure that accepts only valid candidates may nevertheless fail to find a solution. The paper also acknowledges that conventional planners can be much more efficient when a problem is already formally specified. Building a composite system therefore requires a reason for each component.

Self-criticism, progress and the meaning of an explanation

In a study published at ICLR 2025, Kaya Stechly, Valmeekam and Kambhampati tested GPT-4 on arithmetic puzzles, graph colouring and formal planning. Asking the model to critique its own answers often made performance worse in these experiments; correct external verification brought substantial gains. Even repeated attempts with a sound verifier retained much of the benefit of more elaborate feedback. The finding concerns these measurable tasks and configurations, rather than every possible use of linguistic revision.

The group also documented progress. Its 2025 TMLR evaluation, by Valmeekam, Stechly, Atharva Gundawar and Kambhampati, reported substantial accuracy improvements from o1-preview and o1-mini over the earlier language models tested. It examined the additional time and expense involved and showed further gains from combining models with other components. An adequate comparison must therefore consider resources and the system as a whole.

In Beyond Semantics, published in TMLR in 2026, Valmeekam, Vardhan Palod, Stechly, Gundawar and Kambhampati separated correct answers from valid intermediate derivations. Controlled maze experiments showed that training with intermediate sequences transferred from unrelated problems could retain, and sometimes improve, performance. A correct answer did not establish that its accompanying steps formed a faithful derivation. These experiments do not show that every intermediate sequence in every model is meaningless.

A 2026 ICML position paper by Kambhampati and eight collaborators extends that caution to the language used to describe these sequences. Calling them “thinking traces” can encourage readers to treat them as a reliable view of the model’s internal process. The authors acknowledge that generating intermediate content can improve performance. Their objection concerns what users may infer from it about understanding, explanation and trust.

A researcher, teacher and participant in a changing field

Kambhampati served as AAAI president from 2016 to 2018. His 2025 AAAI/EAAI Patrick Henry Winston Outstanding Educator Award recognised teaching materials, mentorship and sustained public engagement with AI and its effects on society. These activities place explanation in another setting: helping students and wider audiences examine the capabilities being claimed for a technology.

The research landscape continues to develop. In independent work published at NAACL 2025, Yilun Hao, Yongchao Chen, Yang Zhang and Chuchu Fan showed how language models and formal verification tools could address complex planning constraints. Such results make the composition of a system central to evaluating its abilities. They also keep the debate open to evidence about which combinations work, for whom and under what conditions.

Across Kambhampati’s career, an enduring question connects the planner, the robot collaborator and the conversational model: what warrants accepting a proposed action? His work has made that question more precise, through implemented systems, experiments and arguments that themselves remain open to scrutiny.

A career in seven moments

1983–1989 · Engineering and planningStudies at IIT Madras and Maryland; completes doctoral research on reusing and modifying plans.
1989–1991 · Stanford to ArizonaWorks at Stanford before joining Arizona State University.
2003 · SapaWith Minh B. Do, publishes a planner addressing time, cost and resource constraints.
2016–2018 · Human-aware AIServes as AAAI president; co-authors the 2017 model-reconciliation approach to explanation.
2023 · Evaluating and modelling actionsContributes to PlanBench, planning evaluations and language-assisted construction of explicit domain models.
2024–2025 · Verification and educationDevelops LLM-Modulo with collaborators, studies self-verification and receives the AAAI/EAAI educator award.
2026 · Answers and intermediate stepsCo-authors experiments and a position paper questioning how intermediate model outputs should be interpreted.