Kim, Brade, Donahue, Huang et al. (CMU/MIT/UCSD/KAIST), 2026

A Design Space for
Live Music Agents

184 systems, 40 years, 3 fields that barely talk to each other — distilled into one map of 31 dimensions and 165 codes for systems that listen and respond to musicians in real time.

Prerequisites: Basic ML intuition + curiosity about music
10
Chapters
5+
Simulations

Chapter 0: The Problem

Imagine you build an AI that jams with a jazz pianist. It listens to her melody and fills in chords underneath, in real time, on stage. You are proud of it. Then you go looking for the literature so you can compare your system to others — and the ground falls away.

An HCI paper describes a "co-creative partner" and spends 12 pages on user experience but never says what model it runs. An AI paper describes a transformer fine-tuned with reinforcement learning and reports millisecond latencies but never asks whether a musician would want it. A computer-music paper talks about "machine listening" and "score following" using vocabulary neither of the other two communities recognizes.

The core problem: Live music agents sit at the intersection of three fields — HCI, AI, and Computer Music — that grew up in parallel. They study the same systems with different priorities and incompatible vocabularies. Without a shared map, every team re-invents, mis-compares, and overlooks. Your jazz jammer might be technically brilliant and socially useless, and you'd have no framework to even notice.

A concrete failure of fragmentation

The paper points to a measurable symptom. Fewer than 10% of generative-audio papers discuss the negative ethical impacts of the systems they propose. Meanwhile musicians repeatedly say they want controllability, transparency, and personalization — and most AI music systems ship with no interaction component at all.

That is the fingerprint of siloed fields: the AI community optimizes a loss function, the HCI community studies a user need, and the two never meet inside one system. Adjacent creative domains — visual art, voice acting, motion capture — have already paid for this with economic disruption and ethics controversies. Music shows the same warning signs.

What "live", "music", and "agent" actually mean here

Before mapping anything, the authors pin down their scope with three crisp tests. A system is in scope only if it satisfies all three:

Notice what this excludes: a real-time synth that just makes sound (not an agent) and an offline music model (not live). The intersection of all three is exactly the slippery, interdisciplinary zone nobody had mapped.

Why a map, not a list

You might think the fix is a big spreadsheet of every system. But a list tells you what exists; it doesn't tell you what's possible. The authors want something that does two jobs at once: look backward (organize 40 years of systems into comparable terms) and look forward (reveal empty regions where nobody has built yet). That two-faced tool has a name in HCI — a design space.

Which system would be out of scope for this paper?

Chapter 1: The Key Insight

The central move of this paper is not a model or an algorithm. It is a structure: a design space organized into four aspects, each containing several dimensions, each dimension offering a menu of codes.

The insight: Any live music agent — past, present, or imagined — can be described as a single choice from each of 31 dimensions. The four aspects answer four different questions: why & for whom (Usage Context), how it feels to use (Interaction), how it's built (Technology), and what it does to the world (Ecosystem). Together they give the three fields a shared vocabulary.

The vocabulary: aspect → dimension → code

This three-level hierarchy is the whole grammar of the paper. Define each level once, with a plain analogy, and the rest of the lesson becomes easy:

ASPECT
A broad question about the system. There are exactly 4. Analogy: chapters of a profile.
↓ each aspect contains
DIMENSION
A fundamental design decision you must make. There are 31. Analogy: a question on a form.
↓ each dimension offers
CODE
One allowed answer to that question. There are 165. Analogy: a checkbox you tick.

Worked example: the Use Purpose dimension (under the Usage Context aspect) has codes like live performance, composition, recreation, skill acquisition, art installation. A given system ticks one or more of these. Your jazz jammer ticks live performance.

The four aspects in one breath

AspectQuestion it answersOwned mostly by
Usage ContextWhy use it, for whom, in what setting, who plays what role?HCI
InteractionHow does it feel to use — interface, control, initiative, timing?HCI + Computer Music
TechnologyHow is it built — model, learning, inference, infrastructure?AI
EcosystemWhat does it do to culture, law, economy, and society?everyone — and nobody

That last row is a spoiler for Chapter 8. The Ecosystem aspect is the one almost everyone skips — and the gap is the paper's most important finding.

Why a design space is two tools in one

The same structure does retrospective and prospective work:

The authors even release the annotated corpus as a living, browsable artifact so the community can keep filling and extending the map.

In the paper's grammar, what is a "code"?

Chapter 2: What Is a Live Music Agent

Before we can map the field, we need to feel the four-decade arc that produced it. The history is short to tell and explains why the three fields drifted apart.

A 40-year lineage in five beats

Collins' definition (2006), still load-bearing: live music agents are "autonomous agents for interactive music, which can at a minimum operate independently of composer intervention during performance." The word "autonomous" is the dividing line between an instrument (you drive it) and an agent (it decides).

The reactive-to-proactive shift

The historical trajectory is a slide along one axis: from reactive accompanists (wait for the human, then respond) toward proactive creative partners (initiate ideas, surprise the human, even lead). Early score followers were purely reactive. Modern systems like jam_bot can take the lead and hand it back. This single shift reorganizes almost every dimension we'll meet.

Three landmark systems you'll see throughout

The paper anchors each aspect to a case study. Meet them now; we'll return to each in its home chapter.

SystemOne-linerAnchors which aspect
jam_botFree-improv partner built with keyboardist Jordan Rudess; trades bars, leads or accompaniesUsage Context
ShimonRobot marimba player with four arms and an expressive head that signals turn-taking with eye contactInteraction
ReaLchordsTransformer + RL that predicts next chords from your melody, in a web app, recovering from your mistakesTechnology
The reactive ↔ proactive timeline
Step through the lineage. Watch systems slide from "reactive" to "proactive".
What single property separates an "agent" from an "instrument" in Collins' definition?

Chapter 3: How the Map Was Built

A design space is only trustworthy if its method is. This is a survey paper, so its "experiment" is the coding methodology. Understanding it tells you exactly how much to trust the percentages you'll see in later chapters.

The data: 184 systems from two worlds

Music today lives as much on YouTube as in proceedings, so the authors deliberately mixed scholarship with practice. Trace the funnel:

731
candidate papers retrieved with Boolean queries over three concept clusters: live, music, agents
↓ six researchers filter by the 3-part scope
117
filtered papers + 36 systems from 34 hand-curated "key papers" = 153 systems from literature
+ YouTube search ("AI jam session", "improvisation with AI")
31
online videos of performances, demos, personal recordings
↓ combine
184
total systems → coded into 31 dimensions / 165 codes

The analysis: inductive coding, not a pre-baked taxonomy

The dimensions were not invented up front — they emerged. The process is worth narrating because it's how every good design space is born:

Tell-tale of the inductive process: they started with five aspects (Task, User, Interaction, Technology, Ecosystem) and merged Task + User into "Usage Context" because so many codes spanned both. The map changed shape under the data — a good sign it's describing reality, not a prior belief.

Why "multiple codes per dimension" matters: codes are mostly mutually exclusive, but a system can tick several within one dimension (jam_bot is both lead and accompanist). This is why percentages within a dimension can sum to more than 100%. If you forget this, the numbers look broken; remember it, and they describe genuine multifunctionality.

How much should you trust it? The reliability number

Coding is subjective, so they measured agreement. About 10% of the corpus (N=19) was double-coded by two annotators, yielding 89.77% inter-annotator agreement. Let's interpret that honestly: ~90% means roughly nine in ten coding decisions were independently reproduced. High enough to trust the big trends; not so high that you should over-read a single 2% code.

What degrades when a system is "N/A"

Not every paper engages every aspect. RAVE — hugely adopted by musicians — is a purely technical contribution; it never discusses Usage Context or Ecosystem. So those dimensions were marked N/A for RAVE and excluded from the denominator for those dimensions. This is the data-flow detail that makes the percentages fair: "85% of systems target musicians" means 85% of systems that addressed Target User at all, not 85% of all 184.

Why can code percentages within a single dimension sum to more than 100%?

Chapter 4: Usage Context — Why & For Whom

Start with the most human aspect. Before any code or model, a designer decides what the activity is for and who is in the room. Usage Context captures the situational frame of the collaboration through eight dimensions.

The eight dimensions, with the headline codes

DimensionTop codes (share of relevant systems)
Use Purposelive performance 80%, composition 26%, recreation 7%, skill acquisition 4%, art installation 3%
Target Usermusicians 85%, novices 17%, audience 7%
Musical Contextgeneral/non-specific 38%, electronic 23%, new music 18%, jazz 10%, classical 10%
Value Emphasiscontrol 39%, coherence 38%, novelty 37%, diversity 28%, personalization 24%
Outcome Emphasisexploration 52%, empowerment 35%, engagement 27%, expression 27%, immersion 10%
User Rolelead 67%, manipulator 21%, non-music performer 15%, conductor 8%
Agent Roleaccompanist 32%, mapper 27%, lead 26%, remixer 24%, controller 11%
Participant Topology1:1 75%, N:1 18%, N:N 10%, 1:N 4%, agent-only 3%
Read the roles together. Users lead 67% of the time; agents accompany 32% and only lead 26%. That asymmetry is the whole field in one number: humans drive, machines support. The reactive-to-proactive shift from Chapter 2 is exactly the slow climb of "agent leads" upward.

User Role and Agent Role are coupled — see it live

These two dimensions aren't independent. When the user is lead, the agent is usually accompanist. When they trade bars, both flip to lead in turn. The interactive below lets you set who leads and watch the natural complementary role light up.

Role complementarity (User ↔ Agent)
User: Lead  |  Agent: Accompanist

Case study — jam_bot: a digital mirror

jam_bot was built with Grammy-winning keyboardist Jordan Rudess (Target User: musician), for live performance in a 1:1 setting, situated in virtuosic practice with no predetermined score. During a set, Rudess and the agent dynamically swap between lead and accompanist.

The design foregrounds two Value Emphasis codes. Control: Rudess needed reliable handles — pressing a root note in the lowest register signals the model to rephrase his recent gestures. Personalization: he wanted it to "feel like a version of myself," realized by fine-tuning the model on MIDI from his own practice sessions. The Outcome Emphasis surprised even him — playing against a model of himself produced reflection: "an analytical look at how I actually think and play."

Concept → realization: "Personalization" isn't a vibe; it's a data pipeline. User MIDI recordings → fine-tuning dataset → updated model weights → an agent whose distribution over next-notes matches the player's idioms. Remove that pipeline and "feels like me" collapses to "feels like the average of the training set."
Across the corpus, the dominant pairing of roles is:

Chapter 5: Interaction — How It Feels to Use

Usage Context told us why. Interaction tells us how it feels in the moment — the functional specification from the user's side. This is where HCI and Computer Music overlap, and where latency stops being a number and becomes an experience.

The dimensions that shape the feel

DimensionWhat it controlsHeadline codes
I/O Modalityraw data in/outaudio 39%/58%, MIDI/symbolic 27%/37%, control 26%/9%, gesture 22%/2%
Temporal Structurehow contributions overlap in timedense parallel, sparse parallel, turn-taking (5%), hybrid, unstructured
Data Alignmenthow data flows between participantscontinuous stream 70%, background triggers, periodic
Control Modehow the user steersimplicit 65%, explicit, none 4%
Control Scopewhat the user can steermaterial 73%, global style 38%, agent behavior 21%, layout 17%
System Initiativewho starts the exchangereactive 60%, mixed-initiative 37%, proactive 3%
Agency Framinghow the system is describedtool vs partner vs hybrid
The two numbers to internalize: control is implicit 65% of the time (the system reads your normal playing rather than asking for commands), and initiative is reactive 60% of the time. Together they paint the median agent as a quiet shadow: it watches what you do and answers, rarely interrupting. Proactive systems — ones that surprise you — are just 3%. That scarcity is a design frontier.

Implicit vs explicit control — the same gesture, two readings

Under implicit control, your musical activity itself is the control signal: you change key, the system modulates with you; you repeat a riff, it generates new material. Under explicit control, you operate a separate handle — a knob for diversity, a sketched melodic contour, a button. Implicit feels like jamming with a musician; explicit feels like driving an instrument. Most systems pick implicit because it preserves musical flow.

Case study — Shimon: turn-taking you can see

Shimon is an embodied robot marimba player (Interface: embodied agent + conventional instrument) with four mallet-wielding arms and an expressive head. It listens to a human's melody, harmony, and rhythm via a MIDI listener (Input Modality: symbolic), and responds by striking the marimba (Output Modality: audio) and moving its head (Output Modality: gesture/visual).

The head is a protocol. Shimon's eye contact is its turn-taking signal: looking toward the human means "you lead"; turning toward the marimba means "I'll take the next phrase." Head bobs provide a beat to follow. This is mixed-initiative + partner framing realized in hardware — the social-communication channel that pure-software agents lack. Strip away the head and the same notes become confusing: you'd never know whose turn it is.

This is the Interaction aspect's deepest lesson: coordination is itself a design problem. Two great soloists who can't signal turn-taking produce a train wreck. Shimon solves it not with a better model but with a better interface for negotiating initiative.

Turn-taking negotiation (Shimon's head as protocol)
Head faces human → human leads.
Shimon's expressive head primarily serves which design purpose?

Chapter 6: Technology — How It's Built

Now the AI side. Technology covers the computational backbone, the learning paradigm, the inference objective, and the brutal real-time constraints that make live music harder than offline generation.

The technical dimensions

DimensionHeadline codes (share)
Modelstochastic process 33%, task-specific DNN 28%, classical ML 25%, generative-AI transformer 6%
Learning Algorithmsupervised 53%, unsupervised, self-supervised, reinforcement learning 5%
Inference Objectiveunimodal generation 52%, classification 24%, cross-modal generation 20%, regression 14%, retrieval 8%
Adaptationnone 64%, offline 25%, online 11%, continual 3%
Technical Emphasislatency 85%, efficiency 27%, error handling 12%, tempo adaptability
InfrastructureMax/MSP, Pure Data, ChucK 62%; general PLs 40%; PyTorch/TF/Magenta
Runtime Requirementscommodity machine 77%, dedicated hardware, cloud APIs, A100/TPU
Integrationbespoke setup 44%, tool-integrated (DAW/VST) 17%, standalone app, source-only
The 85% that explains everything: latency is by far the most-cited technical emphasis. Live music has a hard real-time budget — humans notice control latency above roughly 100 ms. That single constraint is why only 6% use big generative transformers: those models love dedicated GPUs and long inference loops, exactly what the live setting forbids. The field is waiting for model efficiency and commodity hardware to cross.

Why generative AI is "late" to live music

This is the chapter's most important engineering story. Offline text-to-music exploded; live music agents barely use transformers (6%). Why the lag? Two constraints collide:

The escape routes the field is exploring: smaller models, latent representations (encode audio into a compact space, generate there, decode), on-device inference, model compilation, anticipation (predict the human's next move so you've already started computing), truncated context, and distillation. Each trades a little quality for the latency budget.

Case study — ReaLchords: RL for a chord-bot that forgives mistakes

ReaLchords generates chords in real time from your melody (Inference Objective: unimodal generation — predict the next chord token from past melody + chord tokens, both symbolic). Trace its training data flow, because it's a beautiful stack of three learning algorithms:

1. SL
Pretrain the transformer with maximum-likelihood (supervised) on melody+chord sequences → it learns "what chords usually follow".
↓ but MLE-only models are brittle to surprises
2. RL
Fine-tune with reinforcement learning. Reward = musical coherence judged by self-supervised models. Exposing the model to its own predictions (and mistakes) teaches recovery.
↓ plus a teacher that can see the future
3. KD
KL-divergence distillation from a teacher with access to future melody tokens → the student learns to anticipate, sharpening low-latency synchronization.
Concept → realization (why RL here): a supervised model trained only on clean human data has never seen its own bad outputs, so when you play a wrong note it spirals. RL fine-tuning does see the model's own trajectory and rewards getting back on track. That is precisely the error handling Technical Emphasis — and it's why ReaLchords recovers from transpositions and stylistic deviations that break MLE models. Adaptation is offline (finetune before performance); at runtime, inference runs on a remote server so the web client stays light on a commodity machine.
Why do only ~6% of live music agents use large generative-AI transformers?

Chapter 7: The Design Space in Your Hands

Here is the payoff — the paper's Figure 1/Figure 8 reconstructed as something you can operate. A design space is two tools: an analytic lens for existing systems, and a generative canvas for new ones. This widget is both.

How to use it: click codes to build a system profile across six representative dimensions. Switch the mode to Analyze to snap to one of the three case studies (jam_bot, Shimon, ReaLchords) and see exactly which codes each ticks. Switch to Generate to free-form combine codes — and watch the readout flag combinations the survey found empty (research opportunities). This is the whole paper made tactile.
Interactive Design Space — map a system, or invent one
Showing jam_bot. Each ring is a dimension; the lit wedge is its code.

In Analyze mode you can read the paper's case studies the way the authors intended: jam_bot lights up live performance / virtuosic / 1:1 / implicit control / mixed initiative / offline adaptation; Shimon lights turn-taking / partner / mixed initiative; ReaLchords lights unimodal generation / RL / latency+error-handling / standalone. Side by side, the contrasts the paper describes become visible at a glance — that is the analytic power of a design space.

In Generate mode, try building "an agent that listens to music and responds with live code." You'll find no case study occupies it — the survey explicitly names this as an unbuilt cell. That empty wedge is the generative power: the map shows you where to build next.

The same map, in code

A design space is, computationally, just a structured record per system. Here is the core insight as runnable Python — encode systems as code-vectors, then both compare them (analytic) and find empty combinations (generative):

python
from itertools import product

# A design space = dimensions, each with a menu of codes.
SPACE = {
    "use_purpose":   ["live_perf", "composition", "recreation", "skill", "installation"],
    "control_mode":  ["implicit", "explicit", "none"],
    "initiative":    ["reactive", "mixed", "proactive"],
    "model":         ["stochastic", "classical_ml", "task_dnn", "generative_ai"],
    "adaptation":    ["none", "offline", "online", "continual"],
}

# Each surveyed system = a dict of chosen codes (sets allow multifunctionality).
corpus = {
    "jam_bot":     {"use_purpose": {"live_perf"}, "control_mode": {"implicit"},
                    "initiative": {"mixed"}, "model": {"generative_ai"},
                    "adaptation": {"offline"}},
    "shimon":      {"use_purpose": {"live_perf"}, "control_mode": {"implicit"},
                    "initiative": {"mixed"}, "model": {"task_dnn"},
                    "adaptation": {"none"}},
    "realchords":  {"use_purpose": {"live_perf"}, "control_mode": {"implicit"},
                    "initiative": {"reactive"}, "model": {"generative_ai"},
                    "adaptation": {"offline"}},
}

# --- ANALYTIC USE: how different are two systems? (Hamming over dimensions) ---
def distance(a, b):
    return sum(0 if a[d] & b[d] else 1 for d in SPACE)   # & = shared code?

print("jam_bot vs realchords:", distance(corpus["jam_bot"], corpus["realchords"]))
# differs only on 'initiative' (mixed vs reactive) -> distance 1

# --- GENERATIVE USE: which full combinations has NOBODY built? ---
seen = {tuple(sorted(next(iter(s[d])) for d in SPACE)) for s in corpus.values()}
all_cells = product(*SPACE.values())                     # cartesian product
empty = [c for c in all_cells if tuple(sorted(c)) not in seen]
print(f"{len(empty)} unexplored design cells (research opportunities)")
# e.g. ('continual', 'generative_ai', 'live_perf', 'proactive', 'explicit')
#  -> a proactive, continually-adapting generative agent: largely unbuilt

The two halves of the design space fall straight out of the data structure: distance() is the analytic lens; enumerating empty cells is the generative canvas. The paper's contribution is choosing the right dimensions and codes so these operations mean something musically.

In "Generate" mode, an empty combination of codes represents:

Chapter 8: Trends & Gaps

A map's payoff is what it reveals when you step back. The authors apply their own design space to surface trends (where everyone clusters) and gaps (where nobody goes). Three findings stand out.

Finding 1 — The Ecosystem aspect is a desert

Reconstruct Figure 10 mentally: bar heights = number of papers addressing each dimension. The Usage Context, Interaction, and Technology bars are tall. The Ecosystem bars — sociocultural factors, policy considerations, economic consequences, musical-societal consequences — are tiny. Only N=10 systems engage economic consequences at all.

The headline gap: the questions that decide whether these systems are good for the world — Who gets authorship credit? Will musicians lose jobs? Whose musical traditions get marginalized in Western-trained models? — are the least studied. The field optimizes capability and ignores consequence. Adjacent domains (art, voice acting) already show what that neglect costs.
Coverage by aspect (the Ecosystem cliff — Figure 10)
Tall bars = well-studied. Notice the Ecosystem dimensions on the right.

Finding 2 — Generative AI is rising, but slowly

The Model dimension over time tells a clean story: rule-based and stochastic methods dominated early; task-specific DNNs surged after the 2012 deep-learning boom; generative-AI transformers are now appearing — but more slowly than in other computing areas, for exactly the latency/compute reasons from Chapter 6. The prediction: generative AI will take over live music agents once model-efficiency and commodity-hardware trend lines cross.

Finding 3 — Personalization is wanted but barely realized

Personalization is a commonly emphasized value (24%), yet most systems use no adaptation (64%) or only offline fine-tuning (25%). Online (11%) and continual (3%) adaptation — agents that keep learning you as you play — are nearly empty cells. The opportunity the authors name: no-code tools that let non-technical musicians fine-tune and version their own agents, plus test-time adaptation and online RLHF brought into the time-critical live loop.

The honest limitations

The paper's single most emphasized gap in the field is:

Chapter 9: Connections & Cheat Sheet

You can now read any live music agent — paper or YouTube demo — through this lens, and you can spot empty cells worth building. Let's lock it in.

The whole design space on one card

AspectDimensions (the questions)
Usage ContextUse Purpose · Target User · Musical Context · Value Emphasis · Outcome Emphasis · User Role · Agent Role · Participant Topology
InteractionI/O Modality · I/O Musical Element · Musical Outcome · Planning · Temporal Structure · Data Alignment · Interface · Control Mode · Control Scope · System Initiative · Agency Framing
TechnologyModel · Learning Algorithm · Inference Objective · Adaptation · Technical Emphasis · Infrastructure · Runtime Requirements · Integration
EcosystemSociocultural Factors · Policy Considerations · Economic Consequences · Musical-societal Consequences

The numbers worth memorizing

StatisticValueWhy it matters
Systems analyzed184 (153 papers + 31 videos)scope of the survey
Dimensions / codes31 / 165the size of the map
Inter-annotator agreement89.77%how much to trust it
User leads / agent accompanies67% / 32%humans drive, machines support
Implicit control / reactive initiative65% / 60%the median agent is a quiet shadow
Latency as technical emphasis85%the <100ms constraint rules everything
Generative-AI models6%the frontier, throttled by latency/compute
No adaptation / continual64% / 3%personalization wanted, barely realized

How this connects to the rest of the site

The mastery test: pick any music-AI demo you've seen. Can you (1) tick its codes across all four aspects, (2) name one dimension it ignores, and (3) propose a neighboring empty cell worth building? If yes, you own this paper.

"As technological barriers dissolve, the absence of a unified framework becomes the critical bottleneck."

— the paper's thesis, in one sentence
A design space is valuable because it is simultaneously: