💬 NLP

LoSemB: Logic-Guided Semantic Bridging for Inductive Tool Retrieval

The paper proposes LoSemB, a novel framework that leverages logic-guided semantic bridging to enable large language models to effectively retrieve unseen tools in inductive settings by mitigating distribution shifts and reducing reliance on similarity-based retrieval without costly retraining.

Original authors: Luyao Zhuang, Qinggang Zhang, Huachi Zhou, Yujing Zhang, Xiao Huang

Published 2026-01-22

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Luyao Zhuang, Qinggang Zhang, Huachi Zhou, Yujing Zhang, Xiao Huang

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a super-smart robot assistant (a Large Language Model) that can write code, answer questions, and solve problems. To make it even smarter, you give it a giant toolbox filled with thousands of different tools (like weather APIs, flight booking systems, or recipe databases).

The Problem: The "Too Many Tools" Dilemma
The robot's brain has a limit on how much information it can hold at once. It can't memorize every single tool in the world. So, we need a "librarian" (a retrieval system) to pick the right tools for the job.

Most current librarians work like this:

The Transductive Setting (The "Familiar" World): The librarian only learns about tools that exist today. If you ask for a tool the librarian has never seen before (a "new" tool added to the web yesterday), the librarian gets confused.
The Inductive Setting (The "Real World"): The web changes constantly. New tools appear every day. When the robot encounters a brand-new tool it hasn't seen during its training, the old librarian fails. It tries to guess based on the tool's name or description, but because the new tool is different from everything it knows, it picks the wrong one.

The Paper's Solution: LoSemB (The "Logic-Savvy" Librarian)
The authors, Luyao Zhuang and team, created a new system called LoSemB. Instead of just looking at the words in a tool's description (semantic similarity), LoSemB looks at the logic of how tools are used.

Here is how LoSemB works, using simple analogies:

1. The "Social Network" Analogy (Logical Graph)

Imagine every tool is a person at a huge party.

Old Method: The librarian tries to match a new person based only on their name tag. If the new person is named "Cocktail DB," the librarian might think they are a chef because they know the word "recipe." But maybe they are actually a bartender.
LoSemB's Method: LoSemB looks at who the new person hangs out with.
- If a new tool is often used alongside "Flight Search" and "Payment," LoSemB realizes, "Ah, this must be a travel tool," even if it's never seen this specific tool before.
- It builds a map (a Logical Graph) showing which tools are friends with which other tools based on how humans use them together.

2. The "Bridge" Analogy (Logic-Guided Semantic Bridging)

When a new, unseen tool arrives, it feels like an alien because its "vibe" (data distribution) is different from the tools the librarian knows.

The Shift: The paper calls this a "Large Distribution Shift." It's like trying to fit a square peg into a round hole.
The Bridge: LoSemB builds a bridge. It finds tools the librarian already knows that act like the new tool (because they are used in similar situations). It then "borrows" the logic from those familiar tools and applies it to the new one. This aligns the new tool with the old ones, making it easier to understand.

3. The "Two-Step Filter" (Relational Augmented Retrieval)

Once the bridge is built, LoSemB uses a two-step process to find the right tool:

Step 1: The Logic Prune (The Bouncer): Before even looking at the details, LoSemB uses the "Social Network" map to cut out 90% of the tools that couldn't possibly be right. If you ask for a flight, it ignores all the cooking tools immediately, regardless of what their names say.
Step 2: The Enhanced Match: From the remaining small group, it uses a super-powered search that combines the tool's name and its logical connections to pick the winner.

Why This Matters

The paper tested this on real-world datasets (like ToolBench) where new tools are constantly added.

The Result: When faced with tools the system had never seen before, LoSemB was much better at picking the right ones than previous methods.
The Proof: Even when the training data had some mistakes (noise), LoSemB didn't break down. It stayed robust because it relied on the logic of relationships, not just memorizing specific examples.

In Summary:
LoSemB teaches the robot's librarian to stop just reading the dictionary definitions of tools and start understanding how tools work together in real life. By using these logical relationships as a guide, it can instantly figure out how to use a brand-new tool it has never met before, without needing to go back to school (retrain) for weeks.

Technical Summary: LoSemB – Logic-Guided Semantic Bridging for Inductive Tool Retrieval

1. Problem Statement

The paper addresses the challenge of inductive tool retrieval for Large Language Models (LLMs). While equipping LLMs with external tools is a promising paradigm for real-world tasks, the rapid expansion of web-based tool ecosystems makes it impractical to include all tools within an LLM's limited context window. Consequently, retrieval-based methods are employed to select relevant tools from large repositories.

However, most state-of-the-art methods operate under transductive settings, assuming all tools are observed during training. This deviates from reality, where tools are constantly updated and new ones are added. When applied to unseen tools (tools not encountered during training), existing retrieval-based methods suffer from two critical issues:

Large Distribution Shift: Unseen tools exhibit significant distributional divergence from seen tools due to functional diversity. Retriever models trained on seen tools learn biased representations that fail to capture the true functionality of unseen tools.
Sensitivity of Semantic-Only Retrieval: Traditional methods rely heavily on computing semantic similarity between instructions and tool descriptions. When tool representations suffer from distribution shifts, semantic similarity calculations yield misleading results, leading to retrieval failures.

The authors identify that existing methods lack the ability to efficiently extract and transfer hidden logical information to learn better representations for unseen tools without costly retraining.

2. Methodology: LoSemB Framework

The authors propose LoSemB (Logic-Guided Semantic Bridging), a two-stage framework designed to mine and transfer latent logical information to facilitate inductive tool retrieval. The framework consists of an offline indexing phase and an online retrieval phase.

2.1 Logical Graph Construction

The foundation of LoSemB is a Logical Graph ( $G$ ), a bipartite graph containing instruction nodes ( $Q$ ) and tool nodes ( $T$ ).

Edges: An edge exists between an instruction $q_i$ and a tool $t_j$ if the instruction invokes the tool. This forms an incidence matrix (invoke matrix) $R$ .
Embeddings: Initial semantic embeddings for instructions and tools are generated using a fine-tuned Pre-trained Language Model (PLM).

2.2 Logic-based Embedding Alignment

To address the distribution shift, LoSemB extracts logical features from the graph and transfers them to unseen nodes.

Logical Feature Extraction: The framework employs a multi-layer graph convolution (inspired by LightGCN) to propagate information across the graph. The logical features are defined as the difference between the final graph embeddings (which incorporate structural logical relationships) and the original semantic embeddings. This difference captures the node's functional role within the graph.
Similar Tool Identification: For an unseen tool, the system identifies functionally similar seen tools not by text similarity, but by structured propagation. It encodes the unseen tool's associated instruction, finds top- $J$ similar training instructions, and propagates this activation through the invoke matrix to aggregate scores for candidate training tools. Tools with high activation scores are deemed functionally similar.
Logical Feature Transfer: The system transfers logical features from the identified candidate tools to the unseen tool. It uses a frequency-based weighting strategy (for tools) and similarity-based softmax normalization (for instructions) to integrate the weighted logical features with the unseen tool's original semantic embedding. This aligns the unseen tool's representation with the distribution of seen data.

2.3 Relational Augmented Retrieval

The retrieval process combines logical pruning with enhanced similarity matching:

Logical Pruning: Given a test instruction, the system identifies the top- $I$ most similar training instructions and collects all tools invoked by them. This creates a pruned candidate set ( $T_{logic}$ ), significantly reducing the search space by focusing on tools with demonstrated relevance to similar logical patterns.
Graph-Enhanced Similarity Matching: Within the pruned set, the system retrieves the top- $K$ tools by ranking the cosine similarity between the test instruction's graph-enhanced embedding (which includes transferred logical features) and the candidate tools' graph embeddings. This leverages both semantic content and logical relational patterns.

3. Key Contributions

The paper outlines four main contributions:

Problem Identification: The authors identify and analyze the key challenges of existing tool retrieval methods in inductive settings, specifically the large distribution shift and the sensitivity of semantic-only retrieval.
Logic-based Embedding Alignment: A novel module that addresses distribution shifts by integrating logical features (derived from graph convolution) into the representations of unseen tools.
Relational Augmented Retrieval: A retrieval mechanism that leverages both logical pruning (to reduce search space) and graph-enhanced similarity matching (to overcome semantic-only sensitivity).
Empirical Validation: Extensive experiments demonstrating that LoSemB achieves advanced performance in both inductive and transductive settings without requiring costly retraining.

4. Experimental Results

The authors evaluated LoSemB on three benchmarks: ToolBench (I1, I2, I3), UltraTool, and APIGen.

Retrieval Quality: LoSemB outperformed all baselines (including zero-shot LLMs, training-free methods like BM25/AdaEmbedding, and training-based methods like EmbSim/InstRet/RerankRet) in both transductive and inductive settings.
- In the inductive setting with 30% unseen tools, LoSemB maintained stable performance with minimal relative drops (e.g., ~4.18% drop on ToolBench I3), whereas training-based baselines suffered significant degradation (e.g., >10% drops).
Generation Performance: When used to guide downstream task execution (using GPT-4o-mini), LoSemB achieved the highest Pass Rate and Win Rate compared to baselines, indicating that accurate retrieval directly translates to better task completion.
Ablation Studies: Removing either the Logic-based Embedding Alignment or the Relational Augmented Retrieval components resulted in significant performance degradation, confirming the necessity of both modules.
Distribution Shift Analysis: The study showed that the Logic-based Embedding Alignment reduced the KL divergence between unseen and seen tool embeddings from ~0.18 to ~0.001, effectively aligning the distributions.
Robustness: LoSemB demonstrated superior robustness to label noise (incorrect instruction-tool pairs) compared to baseline methods, suffering minimal performance drops under noisy conditions.

5. Significance and Claims

The paper claims that LoSemB introduces a new paradigm for tool retrieval that moves beyond reliance on purely semantic similarity. By mimicking human cognitive processes—where logical relationships and usage patterns are used to understand unseen tools—the framework effectively bridges the gap between seen and unseen tool distributions.

The authors emphasize that LoSemB achieves high accuracy in inductive settings without costly retraining, making it a practical solution for evolving web tool ecosystems. The work suggests that integrating logical information (via graph structures and relational patterns) is crucial for robust tool learning. The paper concludes that this approach paves the way for future techniques involving sophisticated Graph Neural Networks (GNNs) and complex graph structures to further enhance tool learning.