Gander: An Omni Interaction Agent Beyond Turn-Based Dialogue ── A Cerebellum-Brain framework and streaming Thinker-Talker architecture for full-duplex multimodal interaction

The Problem: Turn-Based Interaction as a Structural Constraint

The dominant paradigm in current voice and multimodal AI systems is turn-based: the user completes an utterance, the model processes it, and a response is returned. Neither side can transmit information while the other is speaking. In practice this means users cannot redirect a query mid-sentence, models cannot surface a clarifying question the moment ambiguity arises, and conversations that involve multiple speakers or overlapping information become difficult to handle.

The authors of this paper treat that structural constraint as the motivating problem. They argue that natural human conversation depends on interruption, backchanneling, and proactive mid-process feedback — none of which fit cleanly into a turn-based interface. For applications in document-heavy workflows, multi-party coordination, or any setting where information arrives faster than one turn at a time, that rigidity carries practical costs.

What Gander Proposes

Gander is presented as an end-to-end model with two principal architectural innovations.

The first is the Cerebellum-Brain collaborative framework. The Cerebellum handles real-time interaction and omni conversational capabilities — processing streaming video, speech, and text as they arrive. The Brain handles complex reasoning and higher-level agentic tasks. The two components communicate continuously through tool calling and an agent orchestration runtime. The authors describe this as a division between reactive, low-latency processing and deliberate, higher-order cognition operating in parallel.

The second is the streaming Thinker-Talker architecture. User inputs and model outputs are flattened into an ordered token stream at the chunk level. This unified representation is intended to support low-latency, continuous interaction without the discrete start-stop boundaries of turn-based processing. The Thinker handles internal reasoning; the Talker generates speech output. Running these in parallel allows the model to begin speaking before reasoning is complete, and to revise output as new input arrives.

Together, the authors claim, these designs allow users to interrupt the model at any point while the model can equally emit intermediate feedback or ask follow-up questions without waiting for the user to finish speaking.

What the Evaluation Showed

The authors report evaluating Gander across four dimensions: conversational ability, omni understanding, interactive capability, and agentic intelligence. The primary evaluation method described is internal human evaluation. External benchmark comparisons are referenced where available, but the systematic evaluation reported in this preprint was conducted within the authoring team.

According to the authors, Gander maintains the natural and expressive spoken dialogue capabilities of state-of-the-art open-source models while achieving competitive performance in omni interaction. The paper also reports that Gander demonstrates robustness in challenging real-world scenarios, specifically citing background noise interference, multi-party interactions, and backchannel communication.

These claims have not yet been independently verified. The evaluation being internal to the authors is a material caveat when assessing the weight of these conclusions. The paper is a preprint and has not undergone peer review.

A Concrete Scenario

The following exchange illustrates the interaction model the authors describe. No specific performance figures are implied; this is a structural illustration of the design intent.

This pattern — user interruption mid-utterance, model redirection, model proactively providing status — is what the Cerebellum-Brain design is intended to support. The Cerebellum would process the real-time redirect; the Brain would coordinate the document review task and surface results through the orchestration runtime.

What Is Not New, and What Remains Unresolved

The separation of fast reactive processing from slower deliberate reasoning is not novel in itself. Prior work on cascaded speech systems, streaming ASR pipelines, and multimodal architectures has explored similar divisions. The authors do not claim to be the first to separate these concerns; the contribution is framed as the specific integration within a single end-to-end framework.

The Thinker-Talker design pattern also has precedent in prior speech-language model research. What Gander adds, according to the authors, is the combination with a Cerebellum-Brain layer and an agentic orchestration runtime in one unified model trained end-to-end.

Several questions remain open from the preprint alone. The boundary conditions between Cerebellum and Brain — when reasoning is routed to which component, and how failures in one propagate to the other — are described at a high level. Quantitative analysis of latency under different load conditions is limited to what internal evaluation captured. Endpoint detection (determining when a user has finished speaking versus pausing mid-thought) is a known hard problem in full-duplex systems; the preprint acknowledges the challenge but does not provide detailed ablation of how the model handles ambiguous cases.

The cluster size of one indicates that no strongly related preprints from the same period were identified in the arXiv corpus. This is consistent with an independent line of work, but it also means external comparative reference points are limited at the time of writing.

Why This Topic Is Attracting Attention Now

Full-duplex interaction, streaming inference, tool calling, and agent orchestration have moved from research curiosities to active engineering priorities between 2024 and 2026. Several factors drive this simultaneously.

On the infrastructure side, inference latency has dropped enough that real-time streaming generation is practical at scale. On the application side, agent frameworks — systems where a model calls external tools, reads documents, and coordinates subtasks — have become the dominant deployment pattern for complex knowledge work. The combination of these two trends creates demand for models that can do both at once: respond in real time while managing multi-step agentic processes in the background.

Backchannel communication, specifically — the short acknowledgments and prompts that signal active listening in human conversation — is gaining attention as a reliability signal. A model that only responds in full turns cannot signal to a user whether it is processing, confused, or waiting for more input. The authors cite Gander's performance on backchannel scenarios as evidence that this gap is being addressed.

The open release of models, code, and data (reflected partly in the 176 GitHub stars shortly after publication) follows a pattern in recent multimodal research of prioritizing reproducibility and community development. Whether the release enables independent replication of the reported results is a question that will be answered over the coming months.

Relevance to Pharmaceutical and Regulatory Practice

Document-intensive workflows in pharmaceutical development and regulatory affairs frequently involve sequential review across multiple contributors — a structure that maps onto the limitations of turn-based AI assistance. Submission review, adverse event narratives, and cross-functional protocol discussions all involve information arriving from multiple directions, sometimes simultaneously.

The practical relevance of these capabilities to regulated environments depends on requirements that are separate from the technical design. Auditability — the ability to reconstruct what the model received, processed, and output at each step — is non-negotiable in regulated submissions. Explainability requirements for AI-assisted regulatory decisions are evolving across jurisdictions. Data privacy constraints, particularly where patient information is involved, impose conditions on where and how streaming multimodal inference can run.

None of these requirements are addressed in the Gander preprint, which is not a limitation of the paper so much as a statement of scope. The authors are reporting a model architecture and its capabilities; compliance fit is a separate evaluation that would need to be conducted by any organization considering deployment.

As a research signal, this preprint is worth tracking. It represents a concrete proposal for how full-duplex agentic interaction can be unified in a single model, with code and data released for external examination. Claims should be treated as such until independent evaluation is published.