Overcoming the Manual Predicate Engineering Bottleneck

36 concepts35 papers

Neuro-symbolic AI combines the perceptual power of deep learning with the structured reasoning of formal logic. To bridge these two paradigms, systems rely on predicates—symbolic abstractions like InContact(Robot, Object) or Clear(Block)—that translate continuous sensory data into discrete logical facts.

However, a fundamental bottleneck limits the autonomy of these systems: predefined predicate dependency. Historically, human experts have had to manually engineer the symbolic vocabulary for every new task. When an agent encounters an unfamiliar environment or a novel concept, its reasoning engine grinds to a halt because it cannot reason about what it cannot explicitly name.

To build truly autonomous agents, researchers are shifting from manual predicate specification to autonomous predicate invention.

The Limitations of Predefined Vocabularies

The reliance on human-engineered symbols is a long-standing challenge. Early hybrid systems, such as the SMART+ tool developed by Botta et al., required human experts to provide both the initial predicate set and parameter bounds, restricting the system's utility when the underlying domain theory was unknown.

This bottleneck persists in modern text and reinforcement learning environments. For instance, the SLATE framework by Chaudhury et al. demonstrated high sample efficiency in text-based games but remained constrained by manual keyword matching and fixed predicate sets. The authors noted that this dependency creates a severe bottleneck when transferring agents to new environments with different vocabularies or state properties. Similarly, the NESTA framework by Swaminathan et al. highlighted that symbolic text agents are inherently limited by the fixed vocabulary extracted during parsing, necessitating automated methods for predicate refinement.

In safety-critical reinforcement learning, frameworks like P4P by Francois-Lavet et al. use automata to enforce safety constraints. Yet, because these constraints rely on a manually defined set of atomic propositions linking states to symbols, the agent's capacity to operate in environments where relevant features are not known a priori is severely restricted.

The Grounding Bottleneck in Sensory Domains

The challenge intensifies when moving from text to raw sensory inputs, such as pixels or continuous physical trajectories. In robotics, translating continuous geometric data into discrete logical terms is traditionally handled via hand-crafted heuristics. Tremblay et al. observed this bottleneck when building hierarchical scene graphs for long-horizon manipulation, pointing to a critical need for self-supervised methods that can ground symbolic relationships directly in physical geometry.

Without autonomous grounding, relational learning models remain blind to raw data. As argued by Hazra et al., many relational reinforcement learning systems cannot extract symbolic concepts directly from pixels, requiring significant domain expertise to initialize their symbolic alphabet. Xu et al. identified a similar bottleneck in continuous domains, demonstrating that while symbolic spatial predicates help predict discontinuous physical events like collisions, relying on humans to define these predicates limits true autonomy. This limitation extends to human behavior modeling; Cao et al. noted that logic-informed frameworks for human action recognition often depend on manually predefined spatial and temporal predicates, limiting the system's ability to autonomously discover high-level concepts from raw trajectory data.

The Intuition of Action-Driven Predicate Discovery

How can an agent discover its own symbols without human intervention? The key intuition is that symbols should be defined by what they allow the agent to do.

A foundational insight by Konidaris et al. showed that the exact symbols an agent needs for high-level planning are not arbitrary; they are mathematically determined by the preconditions and effects of the agent's available actions (or options). By focusing on the changes an action causes in the environment, a system can automatically construct its own Planning Domain Definition Language (PDDL) representations.

When this action-symbol link is broken or poorly engineered, planning fails. In his work on neuro-symbolic skills for bilevel planning, Tom Silver demonstrated that modular units combining high-level logic and low-level execution rely heavily on the quality of the underlying predicates; if the human-provided abstractions are insufficient, the entire planning process collapses. To mitigate this, hybrid frameworks like Neuro-Symbolic Relational Transition Models (NSRTs) developed by Chitnis et al. learn relational transitions from data, though they still frequently depend on an initial expert-provided vocabulary.

To push past this, the SLAP framework by Liu et al. allows agents to discover dynamic "shortcuts" during planning, expanding their conceptual model beyond the initial human-provided vocabulary to represent novel abstract states discovered during learning. Similarly, Sreedharan et al. showed that learning optimistic symbolic models on the fly enables efficient exploration in sparse-reward environments, yet emphasized that requiring human-provided predicates to describe world states remains a major limitation on agent autonomy.

Learning Predicates from Feedback and Demonstrations

Instead of relying purely on mathematical preconditions, recent work leverages environmental feedback, human demonstrations, and large language models (LLMs) to infer missing logical primitives.

Environment Feedback and Language Priors

The PSALM framework by Zhu et al. pairs the common-sense reasoning of LLMs with environment feedback to automatically learn symbolic action semantics, allowing planners to operate without pre-defined action rules. Similarly, the authors of Learning Compositional Behaviors from Demonstration and Language introduced BLADE, which uses LLMs to parse human demonstrations and define the underlying logic of skills, though they note that fully autonomous discovery without any human-specified natural language descriptors remains an open challenge.

Disentangling Ambiguity in Demonstrations

When learning from human demonstrations, agents often face ambiguity—for example, determining whether a teacher moved an object because of its color or its position. Zha et al. addressed this with SERLfD, a framework where the agent learns to "self-explain" its success by identifying which high-level relational predicates actually matter.

For task and motion planning, Huang et al. utilized graph attention networks to automatically infer minimal PDDL domains from a small number of demonstrations, bypassing the need for an exhaustive, predefined predicate library. In imitation learning, the HILO framework by Araki et al. connects high-level reasoning with low-level robot trajectories by learning rules from expert demonstrations, but remains constrained by requiring pre-defined mappings between raw sensory data and high-level predicates. This challenge is also mirrored in temporal task inference; Shah et al. framed task specification as Bayesian inference over Linear Temporal Logic (LTL) formulas from demonstrations, but noted that such systems are bound to a predefined set of atomic propositions. Furthermore, in inverse reinforcement learning, Zhou et al. showed that while Symbolic Reward Machines can efficiently learn task rewards from demonstrations, their success is still fundamentally bound to the availability of well-engineered logical predicates.

Differentiable Logic and Predicate Invention

The ultimate goal is predicate invention: frameworks that actively synthesize entirely new logical concepts during training to describe complex data patterns.

Several recent architectures integrate inductive logic programming (ILP) with neural networks to achieve this. The NeSy-$\pi$ framework by Sha et al. automatically invents the logical concepts required to describe complex visual scenes, eliminating the requirement for human experts to pre-define the available predicate vocabulary. Similarly, the NUDGE framework by Delfosse et al. utilizes neurally guided differentiable logic to allow reinforcement learning agents to autonomously invent new logical concepts during environmental interactions, rather than restricting their reasoning to a human-provided vocabulary.

By combining neural perception with automated symbol synthesis, frameworks like DUA (Mitchener et al.) and ISA (Furelos-Blanco et al.) are successfully abstracting raw sensory data into learned automata and subgoals. As these methods mature, they promise to eliminate the manual predicate engineering bottleneck entirely, paving the way for neuro-symbolic agents that can autonomously perceive, name, and reason about the world.

Key Foundational Work

SMART+: A Multi-Strategy Learning Tool (1993) — contributed 1 concept elements to this direction.
Développement autonome des comportements de base d'un agent (2005) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_predefined_predicate_dependency_110f82a6,
  title = {predefined predicate dependency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/predefined-predicate-dependency/rdr_110f82a664054d08ae13250e11c2f35e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_predefined_predicate_dependency_110f82a6_v1,
  title = {predefined predicate dependency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/predefined-predicate-dependency/rdr_110f82a664054d08ae13250e11c2f35e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake