Request Access

GA scheduled for September 2025

Technology

Product

Pricing

Company

Blog

Apollo-1: Neuro-Symbolic AI

Terms of Use

Privacy Policy

Contact Us

Eval Playground

Apollo-1 Playground

Authors: Ohad Elhelo and Ori Cohen, Co-Founders

<h2>Executive Overview: From Generative AI to Neuro-Symbolic AI</h2>

1. The &#8220;Magic Box in the Sky&#8221; (LLMs)
Generative AI models, commonly known as LLMs, emerged rapidly, fundamentally transforming writing, education, code generation, and numerous other tasks. However, despite their linguistic brilliance, LLMs remain fundamentally statistical, predictive text systems: opaque &#8220;black boxes&#8221; lacking explicit, structured reasoning or direct controllability. While the transformer architecture powering these models will certainly continue to evolve, it inherently lacks the design and structure required to serve as a reliable, decision-making brain for conversational agents. Conversational agents deployed by entities—such as banks, insurers, or retailers—require precise, consistent behaviors, predictable policy adherence, and reliable external interactions, all of which are beyond the inherent capabilities of purely generative models. Thus, employing pure LLMs as the complete &#8220;brain&#8221; behind conversational agents in high-stakes contexts is fundamentally flawed and unrealistic.

2. LLM as a Crucial Component, Not the Brain
LLMs, though remarkably sophisticated in linguistic processing—including parameter extraction, classification, intent detection, accuracy validation, and several other linguistic tasks—are not equivalent to cognitive, decision-making intelligence. Just as a steering wheel or engine alone cannot constitute an entire car, LLMs cannot independently form the entire cognitive core required for conversational agents. Their extraordinary linguistic abilities, while critical, represent just one component of a larger cognitive system. Conflating linguistic fluency with complete cognitive reasoning has led to widespread misconceptions about the true capabilities and purpose of LLMs. In a neuro-symbolic architecture, their primary role is to translate unstructured language into structured symbolic representations.

3. Neuro-Symbolic Reasoning: Combining LLMs and Symbolic Logic
The advent of advanced LLMs (GPT-3 and beyond) represented a breakthrough: these models could now reliably automate many linguistic tasks previously performed exclusively by human agents, such as parameter extraction, intent detection, classification, and accuracy validation. Yet, crucially, LLMs remained incapable of trusted final decision-making. Recognizing this limitation, AUI evolved from a purely symbolic approach to a neuro-symbolic architecture:
<ul>
<li style="font-weight: 400;" aria-level="1">NLP Modules (neural): Perform linguistic processing and symbolic encoding. Their dramatic improvements made it possible, without extensive fine-tuning, to reliably create structured symbolic representations across diverse interactions, use-cases, and domains.</li>
<li style="font-weight: 400;" aria-level="1">Symbolic Modules (logic-driven): Serve as the explicit decision-making core, using these symbolic representations to define conversational states, apply structured logical reasoning, enforce policies embedded in the system prompt, and generate dependable, transparent, structured action plans.</li>
</ul>

4. Apollo-1: The Neuro-Symbolic Foundation Model for Conversational Agents
Apollo-1 embodies this transformative neuro-symbolic architecture. At its heart is the Neuro-Symbolic Reasoner, a structured, logic-driven cognitive core explicitly designed as an alternative to transformers for powering conversational agents. By combining the powerful linguistic capabilities of advanced LLMs (for symbolic encoding) with structured symbolic reasoning (for trusted decision-making), Apollo-1 delivers fluent conversational experiences that crucially include reliable actionability, transparency, auditability, and steerability. 
Apollo-1’s design uniquely serves conversational agents explicitly acting on behalf of entities. Unlike purely open-ended chatbots, entity-oriented conversational agents have structured objectives and explicit responsibilities: adhering to organizational policies, managing regulated interactions, and reliably executing actions aligned with the deploying entity’s goals and constraints. Apollo-1’s neuro-symbolic design allows any new interaction, domain, or tool to be immediately integrated simply through structured definitions inserted via the system prompt, enabling instant generalization and adaptability without retraining.

5. The Symbolic Language of Procedural Agentic Knowledge
Recognizing the necessity of explicit, structured reasoning for effective conversational agents, Augmented Intelligence (AUI) has spent nearly eight years meticulously constructing a comprehensive symbolic language, grounded in millions of structured interactions executed by more than 60,000 human agents on a purpose-built dashboard, where actions were encoded into symbolic elements (user intent, parameters, tool use, action plans, etc.). Crucially, this language is not domain-specific; it captures universal, procedural agentic knowledge—how intents are classified, conversational states defined, and actions logically determined—rather than industry taxonomies. 
Initially validated through human-agent interactions, this symbolic approach enabled seamless handoffs among multiple humans managing a single conversation. Over time, the continuous structured interaction data allowed the symbolic language to evolve and mature, eventually becoming clearly interpretable by both human operators and AI engines. Today it serves as a universal, domain-agnostic ontology shared by all Apollo-1 components—NLP Modules, Symbolic Modules, and the System Prompt—ensuring complete internal coherence.

6. Entity-Oriented Conversational AI: The Most Economically Critical Category
This category of conversational AI—agents acting reliably on behalf of entities—represents a substantial, and perhaps the most economically critical, portion of all potential AI applications. Despite advancements, purely transformer-based agents consistently fall short in these economically-critical scenarios, as reflected in multiple recent evaluation benchmarks detailed below, where Apollo-1 decisively outperformed state-of-the-art generative models by wide margins. The reality remains that people still shop, book flights, buy insurance, order Uber Eats, submit insurance claims, and communicate with organizations predominantly through traditional means—not conversational AI—precisely because purely generative transformers are fundamentally unsuitable as the cognitive core for reliable conversational agents.

7. Continuous Improvement Through Structured Human-in-the-Loop Feedback
Apollo-1’s capabilities dynamically evolve through continuous, structured human feedback, integrated directly into its symbolic core. Operators interact via the Reasoning Panel—a real-time dashboard for granular inspection, targeted feedback, and scenario regeneration at the level of individual inference steps and sub-interactions.
Structured feedback drives ongoing improvement along two distinct dimensions:
<ul>
<li style="font-weight: 400;" aria-level="1">Agent-level improvements: Continuous fine-tuning of NLP Modules for accuracy, along with suggested automated updates to the System Prompt, incrementally improve the reliability and effectiveness of individual conversational agents.</li>
<li style="font-weight: 400;" aria-level="1">Model-wide symbolic enrichment: Operator feedback systematically identifies gaps and nuances, strategically enriching Apollo-1’s symbolic language itself. This ongoing symbolic expansion enables Apollo-1 to reliably handle increasingly sophisticated conversational interactions at scale.</li>
</ul>
This integrated feedback mechanism establishes a self-sustaining “flywheel” effect, steadily widening the performance gap with competing systems. Rooted in years of structured interaction data and continuous symbolic refinement, this mechanism provides Apollo-1 with a durable advantage not easily replicated. Additionally, generational advancements in underlying LLM technology directly translate to enhanced NLP Modules for Apollo-1, making linguistic processing progressively faster, more accurate, and cost-effective over time.
The future of conversational AI lies not in replacing generative models, but in harnessing their linguistic excellence within a robust neuro-symbolic reasoning framework. Apollo-1 represents precisely this innovation, opening the era of conversational agents that not only converse fluently, but reliably act on behalf of entities at scale.

<hr />
<h2>Introduction to Neuro-Symbolic AI</h2>
ChatGPT showed the world that machines can talk. But when a machine must also reliably act—book flights, move money, enforce policy—fluent conversation isn’t enough. Banks can’t audit decisions, airlines can’t guarantee bookings, retailers can’t trust inventory calls. Generative AI remains an unreliable black box in scenarios where conversational agents must act on behalf of entities. Yet these scenarios represent the most commercially significant segment of all potential AI use-cases.
For AI to handle these critical interactions, we must transcend purely generative models and embrace a new architecture: Neuro-Symbolic AI. Apollo-1, our neuro-symbolic foundation model for conversational agents, marks the start of this shift. In recent evaluation tests and benchmarks detailed below, Apollo-1 consistently outperformed state-of-the-art generative models by wide margins, precisely on tasks that require conversational fluency combined with dependable, transparent action.
For decades, AI researchers debated two distinct paths: symbolic AI, with structured logic and explicit reasoning; and neural networks, driven by pattern recognition and statistical learning. Both have strengths, both have critical limitations. Purely symbolic AI struggles with natural, conversational interaction. Purely neural, generative AI falters when trust, reliability, and consistency are non-negotiable. The long-standing goal was to combine the two—leveraging each to offset the other’s weaknesses. With Apollo-1, that vision is now reality: conversational agents that converse fluently and act reliably.

The Limitations of Generative-Only AI
The fundamental shortcomings of generative AI have become increasingly visible (and costly):
<ul>
<li style="font-weight: 400;" aria-level="1">Opaque Reasoning: Generative models act as black boxes, leaving no clear explanation of their decisions. This is unacceptable when auditability and accountability matter.</li>
<li style="font-weight: 400;" aria-level="1">Volatile Outputs: Even minor changes in input can drastically alter responses—a hazard in banking, healthcare, customer service, and beyond.</li>
<li style="font-weight: 400;" aria-level="1">Policy Drift: Generative AI regularly ignores or misinterprets critical instructions, rendering it unsuitable for regulated or high-stakes scenarios.</li>
<li style="font-weight: 400;" aria-level="1">Fragile Tool Calls: API calls and complex interactions often fail, especially in multi-step tasks such as bookings and transactions.</li>
<li style="font-weight: 400;" aria-level="1">Costly Retraining: Corrections require costly, time-consuming retraining on massive datasets.</li>
</ul>
Generative AI alone simply cannot fulfill the promises the world expects from advanced artificial intelligence. Recent academic work underscores how far today’s LLM agents are from that bar. A May 2025 <a href="https://arxiv.org/abs/2505.18878">paper</a>, “CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions”, reports that state-of-the-art agents score only 58% on single-turn tasks and collapse to ≈35% once the dialogue spans multiple turns, leading the authors to highlight a “significant gap between current LLM capabilities and real-world enterprise demands.”1

Neuro-Symbolic AI: Bridging the Gap
Generative models sparked the first wave of conversational AI, tailored primarily to individual users. However, when conversational agents must act reliably on behalf of organizations—such as airlines, banks, or retailers—unpredictability becomes an existential liability. The solution demands a shift from purely neural to Neuro-Symbolic reasoning.
Neuro-Symbolic AI bridges the gap between Generative AI&#8217;s linguistic capabilities and Symbolic AI&#8217;s structured reasoning, unlocking actionable, reliable, transparent, and steerable AI interactions. This second wave moves beyond conversation to dependable execution, enabling conversational agents that work for entities, not just end-users.

<hr />
<h2>Introducing Apollo-1: A Neuro-Symbolic Breakthrough</h2>
Apollo-1 is our neuro-symbolic foundation model for conversational agents. It is designed to power conversational agents deployed by entities across industries and use-cases. 
Conversational agents, by definition, are deployed and managed by organizational entities—businesses, brands, or institutions—that must reliably control and align their behaviors. Unlike purely generative models (LLMs) that inherently produce unpredictable and opaque outputs, Apollo-1’s neuro-symbolic architecture was explicitly designed to deliver the transparency, policy compliance, and steerability required for conversational agents that function on behalf of entities. 
Apollo-1 delivers essential capabilities that purely generative models lack:
<ul>
<li style="font-weight: 400;" aria-level="1">Native Tool Integration: Reliable, structured interactions with complex external APIs and systems.</li>
<li style="font-weight: 400;" aria-level="1">Comprehensive Traceability: Each reasoning step is logged in structured, inspectable form, ensuring auditability and compliance.</li>
<li style="font-weight: 400;" aria-level="1">Granular Steerability: Enables precise, real-time behavioral control through structured instructions, context, and guidelines.</li>
</ul>
Apollo-1’s defining innovation is its Neuro-Symbolic Reasoner, a structured, logic-based decision-making core, designed to replace the transformer as the cognitive center of conversational agents. Unlike transformers—statistical prediction engines that produce probabilistic outputs—the Neuro-Symbolic Reasoner systematically performs state management, logical inference and explicit action planning, ensuring complete transparency, compliance with organizational policies, and predictable reliability. 
At the core of Apollo-1 is a universal, domain-agnostic symbolic language, meticulously developed through millions of structured human-agent interactions. Unlike traditional symbolic methods, AUI’s symbolic language captures procedural agentic knowledge, rather than domain-specific content. This symbolic ontology encodes conversational intents, parameters, constraints, and objectives in structured symbolic form, enabling consistent reasoning independent of specific industries or use-cases. Every critical component within Apollo-1’s architecture leverages this unified symbolic language, including:
<ul>
<li style="font-weight: 400;" aria-level="1">NLP Modules: Translate natural language into structured symbolic representations, identifying conversational elements and procedural roles.</li>
<li style="font-weight: 400;" aria-level="1">Symbolic Modules (Neuro-Symbolic Reasoner): Conduct explicit state management, logical inference, and action planning based on symbolic inputs.</li>
<li style="font-weight: 400;" aria-level="1">System Prompt: Defines explicit organizational policies, conversational guidelines, compliance guardrails, and objectives in structured symbolic form, ensuring agent behaviors align with the entity-defined requirements.</li>
</ul>
Because these components share the same unified symbolic ontology, Apollo-1 maintains internal coherence, transparency, and explicit policy alignment throughout every interaction.

Apollo-1&#8217;s Key Neuro-Symbolic Advantages:
<ul>
<li style="font-weight: 400;" aria-level="1">Transparent Reasoning: All reasoning steps are explicitly logged and easily inspectable, unlike opaque &#8220;black-box&#8221; outputs from generative models.</li>
<li style="font-weight: 400;" aria-level="1">Compliance: Structured symbolic validation ensures real-time policy adherence, auditability, and regulatory compliance, overcoming the unpredictable compliance drift common with generative-only approaches.</li>
<li style="font-weight: 400;" aria-level="1">Precise Steerability: Operators precisely steer agent behavior by injecting structured symbolic instructions, constraints, or guidelines—something difficult to achieve consistently with purely generative models.</li>
<li style="font-weight: 400;" aria-level="1">Reliable Tool Integration: Structured symbolic integrations provide robust, reliable connectivity with external APIs and systems, significantly reducing the API fragility typically experienced by generative-only models.</li>
<li style="font-weight: 400;" aria-level="1">Immediate Adaptability: Updates, new conversational capabilities and policy refinements are integrated seamlessly by modifying structured symbolic definitions in the System Prompt, without requiring model retraining or fine-tuning.</li>
<li style="font-weight: 400;" aria-level="1">Continuous Improvement (&#8220;Flywheel Effect&#8221;): Operator feedback, encoded directly in symbolic form, continuously enriches the ontology, steadily expanding Apollo-1’s reasoning capabilities as complexity and scope grow.</li>
</ul>

Beyond Incremental Improvements
Neuro-Symbolic AI fundamentally transforms conversational agents&#8217; capabilities, marking a decisive leap forward. We now have AI that doesn’t just converse but reliably and transparently reasons, adheres to explicit policies, and executes complex tasks seamlessly. With Neuro-Symbolic AI, we finally overcome the chronic limitations of purely generative systems. We get fluent language, yes, but far more importantly, we achieve:

Generative AI (LLMs)

Neuro-Symbolic AI (Apollo-1)

Capability

Apollo-1 does not merely provide fluent language, it reliably reasons, explicitly follows policies, transparently integrates with external tools, and adapts instantly. By merging linguistic fluency with structured symbolic logic, Apollo-1 unlocks the next wave of conversational AI: one specifically optimized for entity-oriented, high-stakes interactions where consistent reliability, controllability, and transparency are essential. While purely generative LLMs remain ideal for creative or exploratory tasks, neuro-symbolic approaches set a new standard for agents that must reliably act on behalf of organizations.

<h2>Technical Overview</h2>
The Symbolic Language: A Procedural Ontology for Conversational Agents
At the core of Apollo-1 is its meticulously constructed symbolic language: a procedural ontology designed specifically to encode agentic knowledge independently of domain-specific vocabularies or industry taxonomies. Procedural agentic knowledge refers explicitly to how conversational intents are recognized, how conversational states transition, and how actions are logically determined within structured interactions.
Unlike symbolic approaches traditionally reliant upon domain-specific ontologies—structured taxonomies defining concepts unique to particular industries—Apollo-1’s symbolic language is universal and domain-agnostic. It systematically encodes procedural conversational roles, logical relationships, and structured reasoning processes essential to reliably performing tasks across any interaction scenario.
This symbolic language was initially developed and refined over eight years through structured interactions executed by tens of thousands of human agents. Through millions of conversational scenarios, each interaction was explicitly translated into structured symbolic representations—identifying intents, parameters, constraints, objectives, and procedural transitions. Over time, this extensive structured corpus evolved into a robust, generalized ontology of procedural knowledge, establishing the symbolic foundations for conversational reasoning independent of specific domains or industries.
Importantly, this procedural symbolic language provides the structural basis for every critical component within Apollo-1’s neuro-symbolic architecture, including:
<ul>
<li style="font-weight: 400;" aria-level="1">The NLP Modules, responsible for translating natural language inputs into structured symbolic forms.</li>
<li style="font-weight: 400;" aria-level="1">Symbolic Modules, responsible for state management, logical inference, and action planning.</li>
<li style="font-weight: 400;" aria-level="1">System Prompt, responsible for encoding explicit instructions, conversational policies, compliance guardrails, and organizational objectives.</li>
</ul>
 
Conversational agents inherently represent entities—businesses, brands, or institutions. Therefore, the symbolic language explicitly encodes organizational instructions, policies, and guardrails provided through the System Prompt. Because all components share Apollo-1’s unified procedural ontology, instructions and policies encoded within the System Prompt seamlessly integrate with conversational intents and parameters identified by the NLP and Symbolic Modules. This unified symbolic representation enables coherent, transparent, and policy-aligned reasoning at all times, regardless of the specific domain or context. This domain-agnostic symbolic design significantly reduces the complexity, time, and cost typically associated with adapting conversational agents to new contexts.
When new conversational inputs are introduced—regardless of their novelty or complexity—Apollo-1 does not depend upon domain-specific vocabularies or hierarchical taxonomies. Instead, inputs are mapped into the symbolic ontology according to their procedural roles, relationships, and constraints, guided by the linguistic context and the predefined policy instructions embedded within the symbolic framework itself. This allows Apollo-1 to immediately generalize and reason effectively about previously unseen interactions, without the need for training, manual symbol creation or explicit adaptation to new domains.
Apollo-1’s symbolic representations are directly human-readable and interpretable, enabling operators to review, analyze, and provide structured feedback through the Reasoning Panel. This structured, stateful feedback—encoded directly into Apollo-1’s symbolic ontology—plays a critical role in the continuous evolution of the symbolic language itself. As Apollo-1 handles increasing volumes of conversational traffic across diverse domains and more complex interactions, operator feedback systematically identifies gaps and procedural nuances requiring enrichment of the symbolic language.
This continuous operator feedback creates a feedback loop—a &#8216;flywheel&#8217; effect—that progressively enriches the ontology, enabling Apollo-1 to rapidly generalize and robustly handle increasingly complex conversational scenarios. This ongoing iterative cycle ensures the symbolic language grows increasingly expressive and precise, progressively strengthening Apollo-1’s ability to generalize robustly across novel and more sophisticated conversational interactions at scale.

The Neuro-Symbolic Reasoner (Inference-time)

In Apollo-1, a Neuro-Symbolic Reasoner replaces the transformer as the model’s decision-making core. As illustrated in the diagram, each conversational turn passes through two distinct, domain-agnostic stages:
<ol>
<li style="font-weight: 400;" aria-level="1">NLP Modules (Encode) – Symbolic Representation 
The process begins when NLP Modules perform natural language understanding, translating conversational inputs—user prompts along with structured context provided by the system prompt—into Apollo-1’s expressive symbolic language. Inputs such as user intents, identified parameters, contextual constraints, and granular operator-provided guidelines are identified and classified. This symbolic language explicitly encodes procedural conversational knowledge into precise semantic units, ensuring accurate, structured, and logically manipulable representations rather than ambiguous linguistic patterns. This approach enables Apollo-1 to natively generalize to new tasks without retraining or manual symbol-creation.</li>
<li style="font-weight: 400;" aria-level="1">Symbolic Modules (Reason) – Logical Inference and Action Planning 
Next, the Symbolic Modules receive these symbolic representations and use them to explicitly define the current conversational state. Leveraging the rich context from the system prompt, they apply explicit logical reasoning to evaluate relationships among entities, parameters, and constraints. They systematically apply logic rules, policies, and guidelines, generating contextual action plans detailing which actions to take and what context each action requires. Every inference step within the Symbolic Modules is transparently documented in the playground’s Reasoning Panel, enabling operators to inspect reasoning steps, provide immediate feedback, or replay and refine scenarios.</li>
</ol>
External Endpoints – Execution 
The contextual action plans generated by the Symbolic Modules translate directly into clearly defined external actions, such as structured API calls, tool invocations, or conversational outputs. Tool responses and acquired knowledge from these actions flow back into the NLP Modules and are re-encoded symbolically. These updated symbolic representations loop back into the Symbolic Modules, iteratively informing further reasoning and actions until the conversational interaction concludes successfully.

The System Prompt: Precise Control Through Structured Symbolic Instructions
Apollo-1’s adaptability across any domain, policy, or tool hinges upon its unified, structured System Prompt, which explicitly encodes organizational guidelines, compliance constraints, conversational objectives, and detailed tool definitions directly into Apollo-1’s symbolic language. These guidelines, originally provided by the operator, are systematically broken down into symbolic instructions, selectively highlighting the parts relevant to the Neuro-Symbolic Reasoner at inference time based on the current state of the interaction.
Unlike generative-only systems—where prompts merely influence probabilistic outputs—the System Prompt in Apollo-1 structurally defines symbolic constraints and instructions that explicitly guide the Neuro-Symbolic Reasoner.
The System Prompt supports extremely detailed definitions of external tools and APIs, including explicit functional specifications, required parameters, pre- and post-conditions for successful execution, expected outcomes, and clearly defined failure states and recovery mechanisms. This granular symbolic representation ensures Apollo-1 can immediately reason about complex external interactions, manage exceptions, and maintain robust performance even in unexpected scenarios, all without retraining or fine-tuning.
When introducing new conversational tasks or integrating new external tools, operators simply insert structured definitions into the System Prompt. This structured approach allows Apollo-1 to immediately comprehend, reason about, and reliably utilize new tools or workflows. Furthermore, the System Prompt enables precise and predictable steering of agent behavior: operators can easily adjust conversational policies, guardrails, or instructions by modifying their symbolic definitions, immediately influencing Apollo-1’s actions in a transparent and predictable manner.
Because NLP and Symbolic Modules consistently interpret every instruction through this shared symbolic ontology, Apollo-1 inherently maintains internal coherence, transparent reasoning, and precise alignment with entity-defined requirements across all interactions and contexts.

Post-inference: Apollo-1’s Human-in-the-Loop Feedback Flywheel
Apollo-1’s performance is continuously improved through structured human feedback, explicitly integrated into its symbolic core. This capability transforms Apollo-1 from a static to dynamic model, allowing incremental enhancements to its reasoning and behavior over time.
Structured human feedback is provided directly by operators through the Reasoning Panel—an advanced evolution of AUI’s original human-agent dashboard. Unlike earlier approaches relying on a large workforce of AUI-contracted human agents, the Reasoning Panel allows organizational operators themselves to explicitly influence and refine Apollo-1’s performance. 
Each reasoning step displayed within the Reasoning Panel is individually clickable, enabling operators to inspect, analyze, and provide granular structured feedback at the sub-interaction level of individual inference steps. Operators can regenerate any specific reasoning step, instantly testing refinements, policy adjustments, or corrections within isolated segments of the conversation.
Each piece of operator feedback, structured explicitly in Apollo-1’s symbolic language, instantly contributes to either agent-wide or model-wide improvements across three critical dimensions:
<ol>
<li style="font-weight: 400;" aria-level="1">NLP Modules Enhancement: (Agent-wide) 
Continuous fine-tuning of the NLP Modules, improving accuracy in intent recognition, parameter extraction, constraint detection, and classification across future interactions.</li>
<li style="font-weight: 400;" aria-level="1">System Prompt Refinements: (Agent-wide) 
Real-time suggested improvements to the System Prompts, including clearer policies, more precise guardrails, and detailed updates to tool specifications, ensuring future interactions more closely align with organizational needs.</li>
<li style="font-weight: 400;" aria-level="1">Symbolic Language Expansion (Model-Wide Generalization) 
Most strategically critical, operator feedback explicitly enriches Apollo-1’s underlying symbolic language itself. This structured, stateful feedback—encoded directly into Apollo-1’s symbolic ontology—plays a critical role in the continuous evolution of the symbolic language.</li>
</ol>
This third dimension—the ongoing expansion and refinement of Apollo-1’s symbolic language—is the catalyst of Apollo-1’s powerful feedback flywheel. As Apollo-1 handles increasing volumes of conversational traffic across diverse domains and more complex interactions, operator feedback systematically identifies gaps and procedural nuances requiring enrichment of the symbolic language. 
This continuous, structured human-AI collaboration—enabled by the uniquely granular, step-by-step feedback mechanism of the Reasoning Panel—is the essence of Augmented Intelligence (AUI), a cybernetic loop empowering Apollo-1 to improve dynamically with every new deployment, use-case, and interaction.

<h2>Hard Proof Across the Toughest Tests</h2>

Results from Google Flights (vs. Gemini 2.5-Flash) and Amazon Retail (vs. Amazon Rufus) tests reflect real-world, one-take scoring. Each scenario allowed only a single attempt per interaction—zero retries permitted. Scenarios required perfect multi-turn execution to pass: any incorrect step resulted in scenario failure.
<a href="https://www.aui.io/resources/apollo-1-one-shots-83-of-111-live-google-flights-scenarios/">Google Flights evaluation</a>: 111 scenarios (booking seats, adding bags, eco-friendly flights, reroutes, upgrades) using real-time inventory from Google Flights. <a href="https://www.aui.io/resources/apollo-1-one-shots-83-of-111-live-google-flights-scenarios/">Amazon Retail evaluation</a>: 120 scenarios (4K TVs, laptop comparisons, color variants, cart management, gift selection, policy queries) using Amazon.com&#8217;s real-time product catalog. Both tests ensured identical API access, ensuring performance differences resulted solely from reasoning and agentic capabilities.
These comparisons specifically utilized Google Flights and Amazon Retail Inventory because these platforms represent some of the most structured and organized API environments globally, providing a rigorous test for an AI agent&#8217;s ability to reliably interact with complex external tools, especially given Apollo-1&#8217;s lack of internal or privileged access to these systems.
* Final Pass^¹, Pass^², Pass^³, and Pass^⁴ numbers will be published in our official τ-Bench report in July 2025.

<h2>What’s Next</h2>
Apollo-1, the first neuro-symbolic foundation model, is already demonstrating transformative results at scale in pilot programs across leading Fortune 500 organizations in critical sectors. Its modular architecture is designed for easy integration with existing generative-AI-based workflows, allowing companies to smoothly transition without a significant operational disruption.
Improvements in underlying LLM technologies directly enhance Apollo-1’s NLP Modules, while ongoing advancements to its symbolic language and Neuro-Symbolic Reasoner independently drive continuous gains in reasoning, transparency, reliability, and steerability. 
Strategic go-to-market partnership with <a href="https://cloud.google.com/blog/topics/partners/google-cloud-partners-with-aui/">Google; </a>General Availability in September 2025, complete with open APIs, toolkits, and rigorous evaluation methods.

<h2>Conclusion</h2>
As researchers, developers, and organizations everywhere grapple with AI&#8217;s complexities and responsibilities, Neuro-Symbolic AI provides a coherent, principled solution: transparency, reliability, steerability, and adaptability, all built into the core architecture.
Generative AI proved machines can talk and think.
Neuro-Symbolic AI proves they can talk, think, adhere to policies, act reliably, and continuously improve through structured feedback.
<hr />
 
Click <a href="https://playground.aui.io/pass-key/">here</a> to explore Apollo-1’s Eval Playground, where you can experience its capabilities firsthand. Interact with the model and its Reasoning Panel, navigate real-time conversations across multiple evaluation domains, and review live benchmark trajectories (passkey required; request access).

<hr />
Additional Resources:
&nbsp;
<table>
<tbody>
<tr>
<td style="text-align: center;">Test / Benchmark</td>
<td style="text-align: center;">Apollo‑1 </td>
<td style="text-align: center;">Best LLM Agent</td>
<td style="text-align: center;">Δ </td>
</tr>
<tr>
<td> 
 
τ‑Bench‑Airline (toughest public benchmark)*</td>
<td> 
 
90.8–92.5 %</td>
<td> 
 
Claude‑4   60 %</td>
<td> 
 
+30 pt</td>
</tr>
<tr>
<td>Google Flights – 111 live booking chats</td>
<td>83 %</td>
<td>Gemini 2.5‑Flash  22 %</td>
<td>≈ 3.5× better</td>
</tr>
<tr>
<td>Amazon Retail – 120 live shopping chats</td>
<td>90.8 %</td>
<td>Rufus  16.7 %</td>
<td>≈ 4.5 × better</td>
</tr>
</tbody>
</table>

<a href="https://www.aui.io/resources/generative-ai-hits-a-wall-on-%cf%84%e2%80%91bench-apollo%e2%80%911-breaks-through/">Apollo-1 tops the toughest agentic benchmark</a>
<a href="https://www.aui.io/resources/apollo-1-one-shots-83-of-111-live-google-flights-scenarios/">Apollo-1 Vs. Gemini 2-5.Flash</a> (Conversational Flight booking)
<a href="https://www.aui.io/resources/apollo-1-amazon-rufus-comparison/">Apollo-1 Vs. Amazon Rufus</a>

<hr />
<h3>References</h3>
1 Huang, K-H.; Prabhakar, A.; Thorat, O.; Agarwal, D.; Choubey, P.K.; Mao, Y.; Savarese, S.; Xiong, C.; Wu, C-S. CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions. Salesforce AI Research (2025). arXiv:2505.18878.