Micros: Diverse Research Frontiers

• Micros are context-dependent systems used in mixed-initiative conversational workshops, syntax-to-intermediate representation transformers, and sub–millimeter microrobots.
• They emphasize explicit internal representations and local decision-making, ensuring fine-grained control in dialogue management, compile-time processing, and autonomous robotics.
• Despite sharing a name, each domain employs unique technical stacks, evaluation metrics, and application-specific challenges.

“Micros” is not a single research artifact but a context-dependent term used in several technically unrelated literatures. In recent arXiv work, it most prominently denotes the Mixed-Initiative ConveRsatiOnal Systems workshop (“MICROS”), syntax-to-intermediate representation transformers (“micros”) for extensible hybrid embeddings in Scheme/Racket-like languages, and sub–millimeter microrobots called “Micros” that integrate on-board sensing, computation, memory, actuation, and communication (Mele et al., 2021, Bocirnea et al., 9 Sep 2025, Lassiter et al., 29 Mar 2025). The term therefore functions less as a stable concept than as a label reused across conversational information access, programming-language implementation, and microscale robotics.

Usage of “Micros”	Domain	Defining characteristics
MICROS	Conversational systems	Mixed-initiative conversational search and recommendation workshop
micros	Programming languages	Syntax-to-IR transformers enabling extensible hybrid embeddings
Micros	Microrobotics	Fully lithographic, reprogrammable, sub–millimeter robots with on-board systems

1. MICROS in conversational information access

MICROS was launched as MICROS@ECIR2021 to convene researchers and practitioners around the next generation of conversational information access, with a particular emphasis on mixed-initiative behavior. Its motivation was the shift toward smartphones, wearables, and smart-home assistants, which made voice-based, interactive experiences mainstream and exposed challenges not adequately handled by traditional search and recommendation paradigms. In this setting, the workshop defined proactivity as the ability of a conversational assistant “to be proactive, proposing not only answers but also possible interpretations for ambiguous or vague requests” and intervening when the system’s confidence is low about the user’s topic of interest (Mele et al., 2021).

The workshop’s technical agenda centered on ambiguity resolution and intent elicitation, dialogue management and contextual understanding, user modeling and personalization, explainability and presentation, and evaluation. Mixed-initiative interaction was treated as more than a user-interface refinement: it required systems to ask clarifying questions, suggest alternative interpretations, maintain conversation history across turns, resolve references such as pronouns, detect topic shifts, and summarize results conversationally under voice-only or small-screen constraints. Compared with traditional web search, this formulation imposed early precision and contextual continuity; compared with batch recommendation, it emphasized preference elicitation, feedback modeling, and strategic selection of the next question.

Methodologically, MICROS explicitly called for deep learning and reinforcement learning for conversational search, conversational question answering, summarization, explanation, and presentation, together with fairness and balance for inclusive systems. Its scope also included conversational query processing and rewriting, relevance feedback, proactive search and recommendation, multi-modal interfaces, user intent and dialog state tracking, knowledge graph presentation, data creation and curation for conversational search, and evaluation metrics for effectiveness, engagement, and satisfaction. Specialized domains such as health, finance, and travel were highlighted because task constraints and risk profiles intensify the need for reliable mixed-initiative dialogue.

Organizationally, MICROS was the first edition of the Mixed-Initiative ConveRsatiOnal Systems Workshop co-located with ECIR 2021. The organizing team included Ida Mele, Cristina Ioana Muntean, Mohammad Aliannejadi, and Nikos Voskarides. The workshop paper did not introduce formal equations or an algorithmic framework of its own; rather, it set a research agenda in which mixed-initiative, proactive strategies were presented as essential for modern conversational systems.

2. “micros” as syntax-to-IR transformers

In programming-language research, “micros” denotes syntax-to-intermediate representation transformers rather than source-to-source transformers. This formulation was proposed as an alternative to macro embedding in Scheme/Racket-like hosts, where macros expand object-language syntax into host-language syntax and remain inside the macro expander pipeline. The central claim is that micro embedding stops at an intermediate representation, producing a deep embedding suitable for high-performance compile-time functions, and only later shallowly embeds the IR back into source syntax. The result is a hybrid embedding: deep for analyses and transformations, shallow for host interoperation (Bocirnea et al., 9 Sep 2025).

The motivation is primarily compile-time performance and extensibility. Macro-embedded type systems can suffer from expansion-order conflicts with type checking and non-local optimizations, because the expander walks terms outside-in while type checking often requires inside-out information flow. The paper states that worst-case quadratic expansion time can result from repeated local expansion and re-expansion of emitted syntax. Micros address this by elaborating surface syntax into efficient compile-time structs, traversing those structs directly, and enabling constant-time field access and in-place mutation.

The IR is represented by mutable compile-time structs with generic methods. Judgments such as type synthesis and type checking are implemented as Racket generics, and the paper introduces an “extensible generics” pattern in which a judgment is wrapped in a parameter so that extensions can dynamically rebind the implementation. It also describes the “mule pattern,” where a macro wraps an IR value into a dummy syntax object carrying the IR in a syntax property, allowing expansion to stop at the IR without re-entering normal source-to-source expansion. Proposed abstractions include define-micro, expand-micros, extensible generics, and extensible structs.

The empirical evaluation compared cur as a macro-embedding baseline, fowl as a micro-embedding system, and smalltt as a highly optimized non-extensible deep embedding. The paper reports that fowl is roughly $1000\times$ slower than smalltt, while cur is roughly $100\times$ slower than fowl when cur does not run out of memory; it also reports roughly linear scaling for fowl with term size and frequent failure-to-complete for cur on larger benchmarks. This makes micros significant less as a new metalanguage than as an implementation strategy for extensible typed DSLs with better compile-time asymptotics.

The paper is also explicit about limits. Strong syntactic extension and strong judgment-level extension are supported, but strong rule-level semantic extensibility remains incomplete because the evaluated system lacks interposition points on struct constructors. No formal theorems of soundness, completeness, or type preservation are provided; the contribution lies in the extensibility taxonomy, design patterns, and empirical compile-time behavior.

3. “Micros” as sub–millimeter autonomous robots

In robotics, “Micros” refers to optically powered, sub–millimeter microrobots fabricated in commercial 55 nm CMOS and designed to sense, think, act, and compute using on-board systems for computation, sensing, memory, locomotion, and communication. The reported devices are comparable in size to a single-celled paramecium. Their common core electronics footprint is $210 \times 340 \times 50\,\mu\text{m}^3$, with total robot widths of $210\,\mu\text{m}$ or $270\,\mu\text{m}$ depending on photovoltaic cell count (Lassiter et al., 29 Mar 2025).

The architecture integrates photovoltaic power harvesters, an optical receiver for downlink, a processor, instruction and data memory, a temperature sensor, actuator drivers for four electrodes, and wiring for platinum actuator electrodes. The processor is a custom CISC design with 11-bit instructions and robot-specific operations mot, ts, and wav, which respectively drive actuators with a programmed pattern, read temperature into memory or registers, and Manchester-encode a register value by modulating motion. On-board digital state is tightly constrained: instruction memory comprises $32 \times 11$ bits, data memory $16 \times 8$ bits, and the register file $4 \times 8$ bits, for a total of roughly 500 bits.

Operation is governed by a power budget of about $100\,\text{nW}$. The processor leakage was designed below about $15\,\text{nW}$, and typical actuation power is about $100\times$0. Locomotion uses electrokinetic propulsion in fluid, with four-electrode robots exhibiting 14 unique polarity states that produce translations, rotations, and arcs. Reported translation speeds are $100\times$1–$100\times$2, and reported rotation rates are $100\times$3–$100\times$4 deg/s. The paper explicitly places the system in the low-Reynolds-number regime, with $100\times$5.

Sensing is demonstrated through an on-chip subthreshold CMOS temperature sensor with resolution of about $100\times$6 and reported accuracy of about $100\times$7 in a total volume of about $100\times$8. Communication is asymmetric: downlink is optical programming via LED pulses at $100\times$9, while uplink uses actuator-modulated motion rather than a dedicated RF or optical transmitter. Programming is passcode-gated and can target all robots or type-specific subsets. After programming, robots operate autonomously on an internal clock and use local sensor inputs to decide actions.

The fabrication route is fully lithographic and massively parallel. Typical release yields exceed 50%, each chip contains more than 100 robots, and a multi-project wafer run yields about 100 chips, or tens of thousands of robots. Current cost is estimated at under US\$210 \times 340 \times 50\,\mu\text{m}^3$00.01 per robot under volume production. Demonstrations include high-precision temperature telemetry by motion and a two-state thermotaxis controller in which the robot changes motion primitive depending on whether the measured temperature rises or falls.

4. Architectural logic shared across the usages

Although these uses of “Micros” are unrelated institutionally and methodologically, they share an emphasis on explicit internal state and locally executed decision procedures. In MICROS, the system must track conversation history, ambiguity, and user intent over multiple turns rather than respond statelessly. In programming-language micros, surface syntax is converted into an explicit intermediate representation on which judgments and transformations operate directly. In robotic Micros, sensing, memory, and control are placed on the robot itself rather than delegated to an external controller (Mele et al., 2021, Bocirnea et al., 9 Sep 2025, Lassiter et al., 29 Mar 2025).

This suggests a recurring pattern: the term “Micros” is repeatedly attached to systems that reject overly transparent or purely reactive models in favor of explicit representations that can be inspected, transformed, or acted upon locally. In conversational systems, that explicit representation appears as dialogue history, user models, and clarification policies. In language implementation, it appears as mutable struct-based IR and parameterized judgments. In microrobotics, it appears as on-board instruction memory, registers, sensor values, and robot-specific instructions.

A second commonality is that each usage links miniature scale or fine granularity to control rather than mere reduction in size. MICROS addressed voice-first and small-screen conversational interaction; micros in programming languages narrowed the compilation target from host syntax to an internal IR; robotic Micros moved decision-making into a few hundred microns of silicon. A plausible implication is that “Micros” in current research naming often signals not simply smallness, but an attempt to make fine-grained systems computationally self-sufficient.

5. Distinctions, misconceptions, and domain boundaries

A common misconception is that “Micros” denotes a unified research program. The literature instead supports the opposite conclusion. MICROS is a workshop name in conversational IR and recommendation; micros in programming languages are compile-time transformers; Micros in robotics are physical micromachines fabricated in CMOS. Their technical stacks, validation criteria, and application domains are disjoint.

Another misconception is that the workshop acronym MICROS should be read as a software framework or deployed platform. The workshop paper is agenda-setting rather than algorithmically prescriptive: it identifies challenges, methods, datasets, and evaluation priorities, but does not define a formal architecture or benchmark suite of its own. Likewise, the programming-language paper does not treat micros as a general substitute for macros in all settings; the argument is specifically about extensible hybrid embeddings, compile-time passes over IR, and the costs of macro-expander-based workflows. The microrobotics paper, conversely, is not about distributed off-board orchestration, but about integrating computation, sensing, and actuation within a sub–millimeter body.

These distinctions matter because superficial lexical overlap could obscure sharply different research questions. Mixed-initiative conversational systems are evaluated in terms of ambiguity resolution, engagement, satisfaction, and effectiveness. Programming-language micros are evaluated by extensibility and compile-time performance. Robotic Micros are evaluated by size, power budget, speed, sensing accuracy, yield, and autonomous behavior. The same label therefore indexes different notions of agency, representation, and scale.

6. Related nomenclature in adjacent literature

The ambiguity of “Micros” is reinforced by neighboring terms that share the “micro” prefix but denote entirely different systems. “Micro Objects” are relatively small immutable containers for distributed data, designed to make distribution explicit by copying objects to where they are needed and organizing updates through append-only clusters of tokens (Wams et al., 2010). “MicroRes” is a resilience profiling framework for microservices that ranks monitored metrics and computes a resilience index from the dissemination of degradation from system performance metrics to user-aware metrics (Yang et al., 2022). “MicroMegascope” is a macroscopic-tuning-fork atomic force microscopy platform that achieves nanometric topographical imaging with a centimeter-sized oscillator and accelerometer readout (Canale et al., 2018).

These adjacent names are not uses of “Micros” in the narrow sense, but they show that the prefix “micro” remains heavily overloaded in contemporary research nomenclature. For that reason, precise citation context is indispensable. In arXiv-indexed literature, “Micros” may refer to a workshop acronym, a compile-time transformation discipline, or a class of autonomous microrobots; without domain qualification, the term is underspecified.