FLAIRS-40 · Deliberative & Generative AI Submit on EasyChair

Special track at FLAIRS-40 (FLAIRS 2027), the 40th International Florida AI Research Society Conference · St. Pete Beach, FL · May 24–27, 2027

Deliberative and Generative AIHybrids and Foundations

A call for papers on systems that combine state-space search with transformer-based generation, and on how deliberative, Type 2 behavior relates to, and might emerge from, generative, Type 1 computation. Abstracts are due January 22, 2027.

Submit a paper on EasyChair FLAIRS-40 conference site
Conference
FLAIRS-40, May 24–27, 2027
Location
St. Pete Beach, Florida, USA
Proceedings
Florida Online Journals, indexed in DBLP and Scopus
Review
Double-blind, by the track program committee

The call

Deliberative AI, the classical, symbolic tradition of state-space search, proceeds through recursive generate-and-test cycles: candidates are proposed, evaluated, pruned, and refined toward a goal. Generative AI, epitomized by transformer-based large language models, is instead characterized by a weak or trivial test step, producing outputs by sampling from learned conditional probability distributions.

A growing body of work treats these paradigms as complements rather than competitors. Search can structure and constrain LLM generation, and LLMs can serve as flexible components within a broader search process: as generators of successors, as scorers, or as tie-breakers. At the same time, an important question remains largely open: how, if at all, does deliberative, Type 2 behavior emerge from a fundamentally generative, Type 1 system?

Following State-Space Search and Transformer-Based AI Hybridizations (Fisher, Moore, and Roberts, ACS 2026), this special track invites papers that explore, build, evaluate, or theorize about hybrid architectures combining state-space search and transformer-based generation. It equally invites papers on the foundational question of how Type 1 and Type 2 processing relate to and give rise to one another. We welcome systems papers, empirical studies, and position or theoretical contributions.

  • How can LLMs serve as generators or evaluators within classical search algorithms?
  • How can search principles improve or control LLM decoding?
  • What is the theoretical relationship between LLM inference and generic state-space search?
  • By what mechanisms might a generative system come to behave deliberatively?

Topics of interest

We see two ends of a spectrum as especially interesting and welcome work anywhere along it: building working hybrid architectures at one end, and characterizing Type 1 and Type 2 processing at the other.

Hybrid architectures

Search and generation working together

  • Proposer-verifier, generate-and-test and LLM-modulo architectures pairing LLMs with symbolic checkers, solvers or simulators
  • LLM-guided search for planning, routing, program synthesis and mathematical proof
  • Retrieval- and generation-augmented reasoning (RAG and GAR) as search-adjacent architectures
  • Neurosymbolic integration, such as concept nets or commonsense graphs paired with generative reasoning
  • Applications in mathematics, transportation and robotics planning, scholarly and bibliographic reasoning, and interactive narrative

Search inside generation

Where the two meet in a single algorithm

  • Search-structured decoding: beam search, lookahead and Monte Carlo-style rollouts applied to LLM generation
  • Scoring and filtering of generated candidates, including self-assessment, auxiliary discriminative models and tool-based scoring
  • Reinforcement learning that folds evaluative, test-like signal back into generative models
  • Computational cost and efficiency trade-offs in hybrid search-generation systems

Foundations

What the categories mean and how they relate

  • Theoretical and empirical characterizations of Type 1 and Type 2 processing
  • How Type 2 behavior might emerge from a Type 1 substrate, in theory or as demonstrated empirically
  • Evidence for latent planning or lookahead within transformer representations
  • Connections to cognitive architectures and dual-process theories of cognition

These topics are examples, not limits. If your work bears on the combination, unification or comparative analysis of search-based and generation-based AI, it belongs here.

Important dates

  1. Abstract registrationRequired: a full paper cannot be submitted without a registered abstract.
  2. Full and short paper submission
  3. Paper acceptance notification
  4. Poster paper submission2 pages plus references.
  5. Poster notification
  6. Camera-ready papers, author registration and early registration
  7. Conference in St. Pete Beach, Florida

All deadlines are 23:59 Anywhere on Earth (AoE). The FLAIRS-40 important dates page is authoritative if anything changes.

How to submit

  1. Prepare an anonymized PDF

    Use the FLAIRS-40 author template from the FLAIRS-40 call for papers. Review is double-blind, so remove author names, affiliations and identifying self-references.

  2. Register your abstract on EasyChair by January 22

    Go to the FLAIRS-40 EasyChair site, start a new submission, and select the track Deliberative and Generative AI: Hybrids and Foundations. Submissions sent to the main track are not reviewed by this track's committee.

  3. Upload the full paper by February 1

    Add the PDF to your registered submission. Poster papers follow the later March 15 deadline.

Full paperUp to 6 pages plus references. Up to 7 pages in the proceedings, with a 20-minute talk.
Short paperUp to 4 pages plus references. Up to 5 pages in the proceedings, with a 10-minute talk.
Poster paperUp to 2 pages plus references. Presented in the poster session.
  • Submissions must be original and not under review elsewhere during the FLAIRS review period.
  • At least one author of each submission must be available to serve as a reviewer.
  • Each accepted paper needs an author registration, and at least one author must present in person.
  • Full papers that are not accepted may still be accepted as short papers or posters.

The FLAIRS-40 call for papers has the complete submission rules.

The motivating paper

First page of the paper, State-Space Search and Transformer-Based AI Hybridizations

Douglas H. Fisher, Kyle A. Moore, and Jesse T. Roberts. 2026. State-Space Search and Transformer-Based AI Hybridizations. In Proceedings of the Thirteenth Annual Conference on Advances in Cognitive Systems (ACS-2026).

Abstract

The paper surveys hybrid deliberative (≈symbolic) and reflexive (≈generative) AI systems, focusing on the combination of state-space search, a workhorse of deliberative AI systems, and transformer-based LLMs, an important class of generative AI. We also speculate within the dual-process framing, notably on Type 2, deliberative computation emerging from Type 1, reflexive computation; and introducing PLuMCoT (Primed Lookahead using Multiple Chains of Thought).

The paper casts LLM generation as a form of greedy search, shows where LLMs can implement the components of a search process, and closes by asking how Type 2 computation might emerge from Type 1. It frames the track's questions, but prior familiarity with it is not expected of submitters.

Read the paper (PDF, 20 pages)

Camera-ready version, hosted here until the ACS-2026 proceedings are online. © 2026 Cognitive Systems Foundation.

Organizers and contact

Questions about the track, or whether a paper fits? Email any of the organizers.