2601.04221
conf
conf
Cooperative Membership Functions for Multi-Agent Oversight
ICLR 2026 · arXiv 2601.04221
alphabell is a federation of small, autonomous research cells working on agentic engineering, world models, recursive self-improvement, and interpretability. No headquarters. No singular executive. Compute pooled across continents. Publications staged, governed by signed proposal and weighted consensus among long-tenured contributors.
Each cell organises itself around one of four research axes. The interpretability axis cross-cuts: every cell working on a dual-use capability is paired with an interpretability cell that has rolling read-access to its checkpoints and training logs. Publication is staged by axis. The pairing is not optional.
Agents as durable computational entities — not wrappers around language models. Persistent state, learned tool affordances, verifiable execution traces; long-horizon planning under partial observability; multi-agent negotiation; sandboxed self-modification; runtime substrates that expose memory, identity, and resource budgets as first-class primitives.
Predictive generative systems that learn the dynamics of physical, social, and symbolic environments from heterogeneous data. Compositional latent dynamics; counterfactual rollout for planning; embodied simulation for robotics pretraining; the unification of perception and prediction under a single training objective.
Models that propose, evaluate, and incorporate modifications to their own training procedures, architectures, and evaluation criteria. Strict internal protocols around capability evaluations, isolated compute enclaves, and pre-registered stopping conditions for any run that crosses defined thresholds. The lab's most closely held research line.
Cross-cutting axis that supplies tooling and theory to the other three. Mechanistic circuit analysis at frontier scale; scalable oversight for agentic systems; formal verification approaches for learned policies; the trust and disagreement protocols that make pairing actually function.
alphabell does not have a headquarters. Four physical anchors exist — small workspaces in the Bay Area, Colombo, Hong Kong, and Bali — but the bulk of the lab works asynchronously through a signed-proposal protocol layer. Compute is committed by federated nodes and allocated through a hybrid mechanism combining tenure, project signals, and quadratic voting.
Member nodes commit GPU and TPU capacity into a shared scheduler. Allocation is computed weekly by a hybrid algorithm: tenure-weighted priority, cell-emitted project signals, and quadratic voting among active contributors.
alphabell runs on a mix of philanthropic grants, sovereign research partnerships, and revenue from selective licensing of derivative tooling. No equity is issued. No funder may unilaterally redirect a research thread.
Foundational results released openly. Capability advances released with delay plus safety analyses. A small fraction — primarily RSI — held indefinitely behind internal review. Every cell signs a research conduct charter on entry.
The internal index is the canonical record of cell-published reports. Most appear here within hours of release; capability releases appear after the applicable delay window. The full index is at /publications.
Releases, governance proposals, funding tranches, safety events, and the occasional reflection from a cell steward. All news is signed by a long-tenured contributor — and where a cell is named, the cell steward of record at posting time.
The name is a play on the alpha/beta release cycle of software — a deliberate signal that the systems we build are never finished, and that the public posture of "ringing the bell" on a finished AI is one we regard as premature.
That bet has cost us things. We do not move as quickly on flashy demos as a centralized lab does. We pay a coordination tax. We make some choices — about publication, about who gets to call a halt — that look from outside like overcaution. We have made those choices on purpose.
Eight roles open this cycle — research engineering and scientist positions in agentic systems, world models, interpretability, and recursive self-improvement; a postdoctoral fellowship in Singapore; three PhD studentships in Lisbon; software engineering on the federated compute substrate; a research-management role on cooperative oversight; and the Visiting Fellow programme for established researchers from external groups.
Compensation is rate-equalised across countries. The lab does not issue equity. The structural commitments — paired interpretability, modification-under-review, signed-proposal governance — apply equally to every role. Applications go to apply@alphabell.com.