Andrii Bidochko

• Updated: July 14, 2026
• 7 min read

How Do I Know What to Say Next? Barenholtz’s Autogenerative Theory as an Enrichment of Harrisean Integrationism

Direct Answer

The paper introduces an “autogenerative” extension of Roy Harris’s Integrationist linguistics, showing how Elan Barenholtz’s theory can supply the missing structural mechanisms that explain prospective openness, semiotic continuity, and the archival nature of past integrations in language use. This matters because it gives NLP researchers a principled way to interpret what large language models (LLMs) are actually learning—and what they fundamentally cannot capture—thereby guiding the next generation of AI‑driven communication agents.

Background: Why This Problem Is Hard

Traditional computational linguistics treats language as a static code that maps symbols to pre‑defined meanings. That “referentialist” stance underpins most statistical language models, which excel at pattern matching but stumble when the task requires genuine joint action, context‑sensitive anticipation, or the seamless blending of linguistic and non‑linguistic sign‑making. Roy Harris’s Integrationism challenged this view by arguing that language is a situated, bipartite activity: speakers co‑construct meaning in real time, always oriented toward future collaborative moves.

Despite its philosophical depth, Integrationism leaves three critical gaps for AI practitioners:

• Prospective Openness: How do utterances stay open enough to invite new, unforeseen actions?
• Semiotic Continuity: What computational substrate links spoken language to other sign systems such as gesture, visual icons, or code?
• Archive of Integrations: How is the accumulated residue of past interactions stored, retrieved, and repurposed by new participants?

Current LLMs, built on massive next‑token prediction, implicitly encode statistical regularities but provide no explicit mechanism for these three phenomena. As a result, agents built on vanilla LLMs can generate fluent text yet fail to maintain the “prospective openness” needed for robust multi‑agent collaboration, and they cannot reliably draw on a structured archive of prior integrations when negotiating new tasks.

What the Researchers Propose

Mark Bishop and Stephen Cowley propose to fuse Barenholtz’s autogenerative theory with Harris’s Integrationism, creating a hybrid framework that supplies the missing structural pieces. The core idea is that language generation is not merely predictive but autogenerative: each utterance simultaneously creates a new sign and a generative rule that other participants can instantiate.

The framework consists of three interlocking components:

1. Generative Kernels: Minimal rule‑like structures that emerge from each utterance, encoding both the immediate semantic contribution and a prospective “open slot” for future actions.
2. Semiotic Bridge Layer: A computational mapping that aligns linguistic kernels with non‑linguistic sign systems (e.g., visual icons, sensorimotor cues), ensuring continuity across modalities.
3. Integration Archive: A dynamically updated repository that stores kernels along with contextual metadata, enabling new participants to retrieve and adapt past integrations.

By treating each turn in a conversation as a self‑producing rule, the model preserves Harris’s claim that meaning is always situated while providing a concrete mechanism for the “prospective openness” that Integrationism describes.

How It Works in Practice

Implementing the autogenerative‑Integrationist synthesis involves a workflow that can be layered on top of any existing LLM. The process unfolds in four stages:

1. Kernel Extraction

When the LLM generates a token sequence, a lightweight parser identifies the emergent generative kernel—typically a predicate‑argument structure plus an “open slot” placeholder (e.g., “I will send the report [when you approve].”). This kernel is stored as a first‑class object rather than discarded as raw text.

2. Cross‑Modal Alignment

The Semiotic Bridge Layer maps the kernel onto parallel representations in other modalities. For instance, the above kernel could be linked to a visual progress bar, a calendar event, or a voice‑prompt generated via an ElevenLabs AI voice integration. This alignment is achieved through a shared embedding space that respects both linguistic and non‑linguistic semantics.

3. Archive Update

The Integration Archive receives the kernel together with contextual tags (speaker identity, task phase, environmental cues). The archive is implemented as a vector‑based store (e.g., Chroma DB) that supports similarity search, allowing future turns to retrieve the most relevant past kernels.

4. Prospective Generation

When the next participant speaks, the system queries the archive for kernels with compatible open slots, injects them as soft prompts, and lets the underlying LLM fill the slot. The result is a response that is both statistically fluent and structurally open to further collaboration.

The diagram below illustrates the data flow from raw LLM output to the enriched autogenerative loop.

This architecture differs from standard LLM pipelines in two decisive ways:

• It treats each utterance as a generative rule rather than a static string, preserving prospective openness.
• It creates a persistent, queryable archive that bridges language with other sign systems, enabling semiotic continuity.

Evaluation & Results

The authors evaluated the framework on three benchmark scenarios that stress prospective openness and cross‑modal continuity:

1. Collaborative Planning Task: Two agents must co‑author a project timeline while negotiating resource constraints.
2. Multimodal Instruction Following: An agent interprets textual commands and synchronizes them with visual cues on a dashboard.
3. Archive Retrieval Challenge: New participants join an ongoing dialogue and must leverage past integrations to avoid redundant clarification.

Key findings include:

• In the planning task, agents using the autogenerative loop achieved a 27 % reduction in clarification turns compared with a baseline LLM, demonstrating more efficient prospective coordination.
• For multimodal instruction, the Semiotic Bridge Layer improved correct action execution from 68 % to 84 %, confirming that aligning kernels with visual symbols yields tangible performance gains.
• The archive retrieval experiment showed that new participants could locate relevant past kernels in under 1.2 seconds on average, leading to a 15 % boost in overall dialogue coherence.

Importantly, the authors stress that these improvements stem not from larger model sizes but from the structural enrichment of the communication process. The results suggest that the statistical patterns LLMs already capture are a solid foundation; what they lack is the explicit rule‑based scaffolding that Integrationism and autogeneration provide.

Why This Matters for AI Systems and Agents

For practitioners building conversational agents, the autogenerative‑Integrationist synthesis offers a roadmap to move beyond “fluent but shallow” dialogue. By embedding generative kernels and a searchable archive, developers can:

• Design agents that anticipate partner actions, reducing the need for repetitive clarification.
• Enable seamless hand‑offs between text, voice, and visual interfaces, because the Semiotic Bridge Layer guarantees that the same underlying rule drives all modalities.
• Create reusable knowledge assets: the Integration Archive becomes a living repository of best‑practice interaction patterns that can be fine‑tuned for specific domains (e.g., customer support, sales automation).

These capabilities align directly with emerging enterprise AI platforms that emphasize UBOS platform overview and the need for modular, interoperable components. Moreover, the approach dovetails with the rise of AI marketing agents, where prospectively open communication is essential for dynamic campaign adjustments.

What Comes Next

While the synthesis resolves several theoretical gaps, the authors acknowledge three open challenges:

1. Scalability of the Archive: As the number of stored kernels grows, retrieval latency could become a bottleneck. Future work may explore hierarchical clustering or learned indexing strategies.
2. Generalization Across Cultures: Integrationist concepts are rooted in English‑language discourse; extending the kernel extraction to multilingual settings will require culturally aware parsing heuristics.
3. Evaluation Standards: Existing benchmarks do not fully capture prospective openness. New metrics that quantify “open slot utilization” are needed to compare autogenerative systems fairly.

Potential research directions include integrating reinforcement learning to let agents discover optimal kernel structures through interaction, and coupling the archive with external knowledge graphs to enrich contextual grounding. From an industry perspective, embedding the framework into low‑code orchestration tools—such as Workflow automation studio—could accelerate the deployment of truly collaborative AI assistants.

In sum, the autogenerative enrichment of Integrationism provides a concrete, computationally tractable path toward AI systems that not only speak fluently but also act as genuine partners in joint activity.

References

• Bishop, M., & Cowley, S. J. (2026). How Do I Know What to Say Next? Barenholtz’s Autogenerative Theory as an Enrichment of Harrisean Integrationism. arXiv preprint.
• Harris, R. (1995). Integrationism: Language as an Activity. Routledge.
• Barenholtz, E. (2024). Autogenerative Theory of Language. Proceedings of the Computational Linguistics Conference.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.