ACL 2026 · San Diego

01 · The architecture

From one giant model to a system of specialists

A frontier model can be remarkably capable and still be the wrong component for every step. It may be expensive, closed, slow, difficult to personalize, or impossible to place near private and latency-sensitive data. A multi-agent system responds by separating responsibilities: one component plans, another retrieves, another calls tools, another critiques, and a smaller local model may handle a narrow decision close to the user.

This changes the central design problem. The question is no longer simply which model scores highest, but how work should move through the system. Useful specialization depends on a controller that can assign the task, expose only the necessary context, collect evidence, and decide whether the result deserves another pass.

02 · Small agents

Small models become useful when their role is precise

Xuan Wang began with a practical question: can small language models become competitive agents? Their advantages are concrete—privacy, local deployment, lower latency, stable operation at the edge—but their reasoning is uneven. A model that looks strong on a fixed benchmark can still fall apart on sorting, planning, or compositional search.

The projects in this part form a useful sequence. BeyondBench diagnoses reasoning with verifiable problems whose difficulty can change. EffGen then treats efficiency as a routing problem, combining prompt compression, decomposition, memory, and complexity-aware model choice. Debate, Train, Evolve turns multi-agent interaction into better training signals, while VISTA-GYM places agents in tool-use situations where intermediate failures, retries, and feedback are part of the task rather than exceptions.

03 · Coordination

Consensus is not the same as truth

Debate makes multi-agent systems look deliberative, yet a group can converge for the wrong reasons. ConsensAgent was presented as an attempt to reach agreement more efficiently and with less sycophancy. The city-planning and political-debate examples showed why this is difficult: goals are ambiguous, information is incomplete, stakeholders conflict, and a confident majority can erase useful minority evidence.

Agreement therefore needs process-level diagnostics. We should be able to see who changed their mind, what evidence caused the change, how much information moved between agents, and whether the system ignored a dissenting but correct signal. Otherwise, collaboration can become a polished way of repeating the same error.

Yuchen Zhuang extended the control problem to collaboration between closed frontier models and trainable white-box models. Hydra shapes the input by selecting better personal context; BBox-Adapter scores and steers the output; Matryoshka Pilot trains a controller around a black-box generator; and Collab-RAG with AceSearcher applies the same principle to retrieval and search. Even when the largest model cannot be fine-tuned, its context, outputs, and multi-step behavior can still be governed.

04 · Biomedicine

In a real domain, the environment becomes part of the model

Wenqi Shi made the architecture tangible. A biomedical agent may have a language model as its “brain,” but it also needs memory, tools, actions, and environmental feedback. It must retrieve domain knowledge, write and execute code, inspect errors, and work with data that can be private, heterogeneous, and noisy.

The examples covered different pieces of that stack. Biomni and the Google AI co-scientist frame broad biomedical ideation; ChemCrow couples reasoning to chemistry tools; EHRAgent answers complex questions over electronic health records through code generation, execution, debugging, and memory; and MedAgentGym turns code-centric biomedical work into an environment where agent behavior can be trained rather than merely prompted.

05 · Scientific work

Science needs environments, not just plausible plans

Scientific discovery makes the agent loop explicit: plan, act, observe, revise. An agent must choose an experiment, spend a limited budget, decide when evidence is sufficient, and recover when execution fails. Without an environment it can narrate a convincing workflow; with an environment it can test that workflow and learn where it breaks.

The tutorial connected single-loop agents, tree search, population search, tournaments, debate, and human- or trace-constrained workflows to AI co-scientists, molecular optimization, autonomous labs, quantum-computing evaluation, and scientific agent gyms. LabOS pushes that picture toward co-scientists that can see and work alongside people, not only read and write.