In brief

  • A major tradition in early AI treated search as an organizing framework for understanding how a system represents a problem, evaluates possibilities, controls its reasoning, and allocates limited computation.
  • Computer science, cognitive science, and neuroscience later divided these questions among specialized research programs. Those programs developed stronger accounts of individual mechanisms, but rarely tested them together in a shared, task-performing system.
  • Language models can already perform many kinds of tasks. Researchers can therefore test how mechanisms for memory, evaluation, control, and search interact without building a new task-specific architecture for every domain.

What an intelligent system must decide

An AI system trying to repair a program has several ways to proceed. It can inspect another file, propose a different fix, run a test, or reconsider its understanding of the bug. Each step consumes time and computation. Writing code is only part of the problem; the system must also decide which operation is worth attempting, whether the result should change its plan, and when the evidence is strong enough to stop.

At Voaige, we kept encountering versions of this problem both in the literature on memory, confidence, planning, and cognitive control and in our experiments to augment language models. The literature we surveyed offered increasingly precise accounts of individual mechanisms. We had a hunch that these mechanisms could work together, so we explored the history of treating them as components of intelligent systems.

The history was older and less linear than we expected. By the end of the 1980s, researchers had built programs that coordinated specialized reasoning processes, changed the rules they used to decide what to do next, and learned from experience solving problems. Their work raised the questions that guide this account: what did those systems accomplish, why was their progress so difficult to carry across domains, and what has changed since?

Early AI used search as an organizing framework

In 1950, Claude Shannon described a chess-playing computer that considered and evaluated future positions. The full game tree was far too large to examine, so a useful program had to choose where to look and how far to continue. Shannon distinguished a broad search to a fixed depth from a more selective strategy that pursued promising lines further (Shannon, 1950).

Allen Newell, Herbert Simon, and Cliff Shaw carried a related approach into reasoning. Their Logic Theorist searched for proofs using heuristics to select promising symbolic operations (Newell & Simon, 1956). The General Problem Solver, or GPS, then tried to make a problem-solving method reusable across domains. It compared the current state with a goal, chose an operation that could reduce an important difference, and created subgoals when the operation could not yet be applied (Newell, Shaw, & Simon, 1959).

GPS required substantial work before the first search step. Someone had to specify the relevant objects, the available operations, the differences that mattered, and the conditions for success. During the search, the system still had to choose which difference to address and which subproblem to pursue. Early AI researchers called these choices control: the decisions that determine which operation to try, where to direct attention or computation, and when to stop. Search therefore provided an organizing framework for representation, knowledge, and control. GPS separated a task's description from the method used to solve it, but each new task still needed a useful description.

Once programs contained several kinds of knowledge and several possible operations, researchers also had to decide which capability the program should use next.

Early architectures made coordination explicit

By the early 1970s, Newell was asking the same question about psychology. Researchers had accumulated experimental effects and plausible accounts of individual mechanisms, but a collection of such accounts still left the sequence of behavior unexplained. A theory also needed to say how operations were selected and combined while a person performed a task (Newell, 1973).

Newell proposed building complete processing models and requiring the same system to perform many tasks. This imposed a useful discipline: a mechanism introduced to explain one result would have consequences elsewhere, and changing one element of the theory could disrupt its account of another task.

Speech understanding showed why coordination mattered. A noisy sound can fit several words. The surrounding syntax may favor one interpretation while the meaning of the sentence favors another. Committing too early can discard the correct possibility; preserving everything can overwhelm the system.

Hearsay-II, developed at Carnegie Mellon in the 1970s, organized this work around a shared “blackboard.” Specialized programs contributed partial interpretations at different levels, from acoustic segments to phrases, and acted when their knowledge became relevant. If the acoustic evidence weakly supported two words, both could remain available while syntactic and semantic processes supplied additional constraints. The interpretation developed through interactions among these contributions rather than through a fixed pipeline (Erman et al., 1980).

Several processes might be ready to act at once, so a scheduler had to decide which opportunity deserved computation. This was the control problem in concrete form: the system needed rules for choosing among useful next steps. BB1 later represented those decisions on a second blackboard. That addition made the system's method for directing its own reasoning available for inspection and revision, rather than leaving it fixed in the scheduler (Johnson & Hayes-Roth, 1987).

Cognitive architectures extended the same idea beyond a single application. They specified how memory, perception, learning, and action would interact in a reusable design, then tested whether that design could support many tasks. In Soar, the absence of a rule for choosing what to do next created an impasse. The system treated the impasse as a new problem, reasoned through it, and could compile a successful resolution into a rule for later use. Search was therefore not only guided by prior knowledge; experience could change how future searches were controlled (Laird, Newell, & Rosenbloom, 1987). ACT, and later ACT-R, developed a different account of how procedural knowledge, declarative memory, perception, and action jointly produce behavior (Anderson et al., 2004).

Newell's Unified Theories of Cognition placed these projects inside a larger scientific argument. He described the field as having reached the possibility of unified theories, not their completion (Newell, 1990, 1992). By the early 1990s, several substantial programs made integration seem possible enough to build toward, but none became an accepted model of intelligence.

Learning improved search while metareasoning formalized control

The games lineage shows how learning changed the source of the knowledge that guided search. Deep Blue relied on chess knowledge and an evaluation function substantially designed by people. AlphaGo instead trained policy and value networks on game data and self-play, then used what those networks had learned to direct and evaluate tree search. AlphaZero extended the approach across chess, shogi, and Go, while MuZero learned an internal model useful for planning rather than receiving a complete simulator of its environment (Campbell, Hoane, & Hsu, 2002; Silver et al., 2016, 2018; Schrittwieser et al., 2020).

Here, learning means acquiring from data or experience some of the knowledge that a programmer previously had to encode: which possibilities are promising, how a position should be represented, and how progress should be evaluated. From Deep Blue through MuZero, learning supplied more of this guidance. These systems still relied on unusually favorable conditions, including specified actions, repeated trials, and clear outcomes. Open-ended work often requires a system to construct the possibilities itself and determine what evidence would distinguish them.

Metareasoning formalized another part of the same problem. Computation consumes time and resources, so a bounded system must decide whether another step is likely to be worth its cost. Horvitz analyzed reasoning in terms of the value of its results and the resources it consumed, while Russell defined rationality relative to what a program could achieve on a particular machine in a particular environment (Horvitz, 1988; Russell, 1995). This work turned control from a scheduling detail into a central feature of rational behavior.

Cognitive science and neuroscience sharpened the distinction between monitoring—estimating whether current reasoning is reliable or making progress—and control—using that estimate to choose what to do next. Research on confidence, expected payoff, efficacy, and effort helps explain why recognizing a problem is not the same as responding well to it (Shenhav, Botvinick, & Cohen, 2013; Fleming, 2024). A system can detect uncertainty yet spend its remaining effort on an unhelpful step. Improving its confidence estimate and improving the policy that acts on that estimate are different interventions.

Integration continued in working programs

In the 2000s, DARPA's Personalized Assistant that Learns program and its largest project, SRI's Cognitive Assistant that Learns and Organizes (CALO), brought more than 300 researchers at 22 institutions together around a working cognitive assistant. Advances in learning, planning, language, interaction, and decision-making had to function within the same system. CALO demonstrated what coordinated integration could build, along with the organizational effort required to sustain it (SRI International, 2003; DARPA, n.d.).

Executable cognitive architectures sustained one version of the integration project. Soar and ACT-R required proposed mechanisms to work together in programs that performed tasks, while the Common Model of Cognition identified commitments shared across those architectures and related systems (Laird, Lebiere, & Rosenbloom, 2017).

Other programs organized integration around shared scientific questions rather than a single architecture. MIT's Center for Brains, Minds and Machines (CBMM) brought together integrated research programs guided by the goal of a unifying theoretical foundation for intelligence. Berlin's Science of Intelligence studied principles that might apply across biological and artificial systems and used them to guide mechanistic hypotheses and the construction of artificial agents.

CALO, cognitive architectures, CBMM, and Science of Intelligence demonstrate that integration did not disappear after 1990. They also show why it remained unusual. Each program depended on a shared object that different specialists could develop together: an assistant, an executable architecture, a body of comparative evidence, or principles embodied in artifacts. Integration requires someone to maintain that shared object, define the interfaces between ideas, and evaluate whether the combined result improves.

Why these programs remained exceptional

These programs were exceptions to the broader organization of research on intelligence. Neuroscience, for example, produced increasingly detailed accounts of memory, confidence, cognitive control, valuation, and learning, but usually studied them within separate traditions organized around particular mechanisms, tasks, and methods. The field developed stronger explanations of individual functions without establishing a comparably sustained program for explaining how they jointly produce intelligent behavior during complete tasks.

Mario Senden's analysis of 461,316 neuroscience papers published from 1999 to 2023 illustrates this pattern. The study identified 175 research clusters and found extensive citation across domains, showing that ideas and methods circulated between specialties. Yet 64 percent of the clusters were characterized by narrowly scoped mechanistic accounts, while only 17 percent contained work employing overarching theoretical frameworks. Neuroscience was intellectually interconnected, but synthesis into broad theory did not become one of its main organizing projects (Senden, 2026).

Kriegeskorte and Douglas described a related gap in 2018. Cognitive science had decomposed task performance into computational components, while computational neuroscience had modeled the neural implementation of particular component functions. Neither had yet explained how those components operate together to produce cognition and behavior (Kriegeskorte & Douglas, 2018).

Closing that gap requires a different kind of research infrastructure. A memory mechanism contributes to intelligent behavior through what a system retrieves, when it retrieves it, and how that information changes subsequent decisions. Confidence matters through the actions it changes; a control policy matters through the processes it coordinates. Testing these interactions requires working systems, expertise across disciplines, and measures that capture the combined result. Academic fields more readily support experiments that isolate one mechanism and evaluate it according to the standards of an established specialty.

This makes the integration programs described above important for two reasons. They preserved the ambition of studying intelligence as coordinated behavior, and they revealed what that ambition requires: a shared object that specialists can develop together, explicit interfaces between mechanisms, and sustained evaluation of the combined result.

Language models change the starting point

The earlier integration programs depended on a shared working object, but constructing that object was itself a major research project. Modern language models change this starting point. They can already work on many problems expressed through ordinary language, code, and other learned representations, so researchers can investigate how memory, evaluation, and control should interact without first encoding a new symbolic world for each task.

Recent methods expose different decisions within that control problem. An important precursor in recurrent neural networks, Adaptive Computation Time, made the amount of processing input-dependent (Graves, 2016). Chain-of-thought prompting makes intermediate textual steps available for further generation or evaluation, although those steps are not transparent records of a model's internal computation (Wei et al., 2022). Other methods change how a model explores and assesses alternatives: self-consistency generates several paths and aggregates their answers; Tree of Thoughts adds evaluation, lookahead, and backtracking; ReAct lets observations from external actions redirect the reasoning process; and process supervision trains evaluators to judge intermediate work (Wang et al., 2023; Yao et al., 2023a, 2023b; Lightman et al., 2023).

Research on test-time compute has made the consequences measurable. Snell and colleagues found that the best allocation strategy depended on the problem and the underlying model, and that optimized test-time computation could outperform a larger model under particular matched-compute conditions (Snell et al., 2025).

Together, these methods make several aspects of control experimentally accessible. Researchers can vary how long a model reasons, how many alternatives it considers, what evidence it gathers, how it evaluates progress, and when it stops.

Why integration is more tractable now

Because language models already have broad task competence, researchers can test interactions among mechanisms without building a new architecture for every domain. They can manipulate proposals, evidence, evaluators, tools, budgets, and stopping rules at test time. Neuroscience, cognitive science, and computer science now offer richer accounts of monitoring, control, memory, valuation, exploration, and learning. The engineering objective is also concrete: improve the result for the time and computation spent.

That is the effort we are building at Voaige. We bring researchers across neuroscience, cognitive science, and AI together to identify computational principles of cognition and engineer them into modern models. Our organizing question is how a system should coordinate the decisions required for search under limited time and computation across domains: how it represents a problem, constructs candidates, allocates computation, evaluates progress, stops, and learns from the result.

This framework is a working decomposition rather than a final theory of intelligence. Its mechanisms may be learned through experience, constructed during the current attempt, specified by a designer, or produced through a mixture of all three. We evaluate them by whether they improve accuracy and efficiency across tasks, first independently and then in combination. A mechanism that succeeds in isolation may fail once it must share information, computation, and decisions with the rest of the process, so the combined behavior is the relevant test.

Earlier programs supply ideas, practical lessons, and potential collaborators. Modern models supply a more capable experimental starting point. The opportunity is to turn those resources into a cumulative engineering practice in which proposed principles earn their place by improving the combined result.

References

Continue exploring

Research on cognition, search, and more powerful adaptive intelligence.

Explore Voaige research Back to Blog