UniR: Plug-and-Play Modular Reasoning

• The paper introduces a novel plug-and-play modular reasoning framework (UniR) that augments frozen LLM backbones with composable neurosymbolic modules.
• It employs a Neurosymbolic Transition System to integrate symbolic transitions and neural inference, ensuring termination, soundness, and completeness.
• Empirical results show UniR boosts mathematical reasoning pass@1 accuracy by 36% and enables flexible module composition across diverse tasks with minimal retraining.

Plug-and-play modular reasoning for LLMs denotes a design paradigm in which specialized “reasoning modules” can be composed, reused, and hot-swapped atop frozen LLM backbones to augment or guide their reasoning abilities. This modular approach enables robust integration of symbolic and neural components, efficient specialization for complex tasks such as mathematical or programmatic reasoning, and the ability to combine multiple reasoning techniques flexibly at inference time—all without incurring the prohibitive costs of full model retraining or eroding the base model’s generalization. A canonical instantiation is UniR (Universal Reasoner), which demonstrates task-conditional composability at the token level and supports principled modularity through explicit interface contracts and theoretical guarantees (Kim et al., 25 May 2025, Bembenek, 8 Jul 2025).

1. Neurosymbolic Transition Systems: A Principled Model for Modular Reasoning

The foundation for plug-and-play modular reasoning is the Neurosymbolic Transition System (NSTS) model (Bembenek, 8 Jul 2025). An NSTS is defined over pairs $(s, i)$, where $s \in \Sigma$ is a symbolic state and $i \in I$ is an “intuition” data structure (e.g., prompt strings, neural embeddings). The core dynamics are:

• Symbolic transitions: $s \rightarrow_s s'$ represent base algorithmic logic (e.g., proof-tree expansion, search step).
• Intuition updates: $i' = \text{combine}(i, \omega(s \rightarrow_s s'))$, where $\omega$ injects developer-supplied neural guidance for a transition.
• Neural “inference”: Optionally, $\text{infer}(i)$ queries an LLM to suggest or bias the next symbolic move.

An NSTS thus advances by explicit lock-step transitions of the symbolic state and corresponding neural intuition, providing a substrate where modules encapsulate transition rules and neural hooks (intuitive evidence, biases, or contextual notes). The NSTS model ensures that symbolic guarantees—termination, soundness, and completeness—hold for the overall system. Every module operates over its own symbolic and intuition spaces, registered through contracts such as $\text{declare\_syntax}$, $\text{declare\_transitions}$, and $\text{declare\_infer\_strategy}$. This modular registration enables plug-and-play extensibility (Bembenek, 8 Jul 2025).

2. UniR: Universal, Composable Reasoning Modules for LLMs

UniR formalizes the plug-and-play reasoning interface as a lightweight, standalone transformer module, $\pi_r(\cdot | x; \phi)$, trained with predefined (e.g., verifiable) rewards using GRPO (Group Relative Policy Optimization) while keeping the LLM backbone frozen (Kim et al., 25 May 2025). UniR decomposes a reward over entire model trajectories,

$r(x, y) \quad (x \in \mathcal{D}, y = y_{1:T}),$

into token-level log-probabilities by optimizing

$$\frac{1}{\beta} r(x, y) = \sum_{t=1}^T \log \pi_r(y_t | x, y_{<t}; \phi).$$

At inference, UniR’s output logits are simply added to the frozen LLM’s logits, enabling immediate synergistic decoding, i.e.,

$$\mathbf{z}_{\text{total}}(x, y_{<t}) = \mathbf{z}_0(x, y_{<t}) + \lambda\, \mathbf{z}_r(x, y_{<t}),$$

with $\lambda$ a tunable scale. This additive architecture supports direct modular composition: multiple independently trained UniR modules ($N$ distinct $\pi^i_r$) are combined by logit sum, yielding multi-objective, token-wise decision-making without further retraining. This strictly plug-and-play approach is agnostic to the backbone and introduces neither communication nor dependency on internal LLM states (Kim et al., 25 May 2025).

3. Modular Registration and Interoperability

Plug-and-play modular reasoning in UniR (and the NSTS paradigm) relies on standardized module interfaces. Each module exposes:

• Local symbolic state shape ($\Sigma_M$), transition rules ($\rightarrow_s^M$)
• Local neural attachment ($\omega_M$), defining how to embed module-specific transitions into the global neural intuition
• Optional custom inference strategies (e.g., $\text{infer}_M$) for distinct LLM prompt patterns

A global engine (NSTS or UniR runtime) composes transition relations and intuition updates from all registered modules. Because each module only manipulates its local symbolic state and neural attachment, modules can be swapped, reordered, or extended with zero changes to the core engine. In practice, this supports hybrid deployment scenarios: mixing SAT solvers with neural program-synthesizers, type-checking modules, or specification checkers within a single unified decision process (Bembenek, 8 Jul 2025).

4. Empirical Results, Transfer, and Practical Characteristics

UniR demonstrates substantial empirical gains across diverse reasoning domains. On mathematical reasoning tasks (GSM8K, MATH-500, AIME24), UniR with a 1B parameter module plus a 3B Llama backbone delivers in-distribution pass@1 accuracy of 63.8% and out-of-distribution averages of 15.0%, outperforming full 3B parameter fine-tuning and LoRA baselines, with an overall average improvement of 36.0% (Kim et al., 25 May 2025).

For machine translation, UniR achieves higher BLEU and CometKiwi scores than baselines, and composition of modules allows joint optimization over multiple reward objectives. UniR is robust to backbone scaling: a module trained on a small backbone (3B) can significantly improve larger backbones (e.g., 14B) with no retraining (“weak-to-strong generalization”). Training cost is drastically reduced, as only the reasoning module is updated. Memory footprint is minimized: batch sizes up to 128 fit on standard 80 GB hardware, whereas LoRA/full-finetuning are more restrictive.

5. Comparison with Other Plug-and-Play Reasoning Frameworks

Other plug-and-play frameworks, such as TART (Bhatia et al., 2023) and Chameleon (Lu et al., 2023), affirm the robustness of modular reasoning for LLMs across settings:

• TART: A decoder-only transformer, trained solely on synthetic logistic regression tasks, can be prepended to arbitrary frozen LLMs and used to perform in-context Bayesian-style inference. TART improves average accuracy in low-shot classification tasks by +18.4 points and transfers across modalities by operating on PCA-reduced LLM embeddings. However, it is limited to (binary/continuous) classification—generalization to higher-order logic and output structures remains non-trivial (Bhatia et al., 2023).
• Chameleon: LLM-guided modular planning orchestrates heterogeneous tool compositions (LLMs, vision models, web tools, program synthesis, tabular heuristics) to realize compositional reasoning pipelines. The planner, itself an LLM, selects and sequences modules as dictated by task and context; modules are executed sequentially with cached input/output updates. Chameleon achieves >86% accuracy on ScienceQA (multi-modal, knowledge-intensive QA) and 98.78% on TabMWP (math over tables), outperforming monolithic LLM prompting and human baselines (Lu et al., 2023).

A trending theme in all systems is abstraction and reusability: once trained, modules are reused across LLMs, benchmarks, and modalities.

6. Theoretical Guarantees and Inheritance of Symbolic Properties

Plug-and-play modular reasoning frameworks inherit critical guarantees from their symbolic substrates by construction. Specifically, an NSTS only allows symbolic transitions valid under the base rules ($\rightarrow_s$). As a result, termination, soundness, and completeness properties are preserved: if the symbolic process is k-bounded and outputs only correct solutions, so does the composed system. The neural reasoning layer never adds spurious transitions nor blocks legal symbolic progress; neural modules serve purely as bias or guidance, subject to the constraints of the symbolic system (Bembenek, 8 Jul 2025). For multi-objective token-level orchestrations (UniR), modular composition via additive logits supports tradeoff exploration without additional learning.

7. Outlook, Limitations, and Extensions

Plug-and-play modular reasoning with LLMs, as realized in UniR and NSTS, addresses key shortcomings in prior ad hoc neurosymbolic integration: lack of strong guarantees, difficulty in synchronizing neural and symbolic information, limited extensibility, and high adaptation cost. However, structural limitations persist:

• Module performance is bounded by the expressiveness of reward decomposition; tasks demanding higher-order logical reasoning or chain-of-thought may require new module architectures (Bembenek, 8 Jul 2025, Bhatia et al., 2023).
• In purely LLM-based planners (Chameleon), context window and planning granularity bottlenecks can arise for large module inventories (Lu et al., 2023).
• NSTS and UniR assume modules are reliable and interaction effects minimal. Real-world tool reliability and inter-module dependencies may challenge these assumptions.

Future work includes dynamic re-planning, on-the-fly module registration, support for heterogeneous data modalities, and further theoretical analysis of multi-module interactions and convergence. The plug-and-play modular approach constitutes a rigorous, extensible substrate for scalable LLM-powered reasoning systems.