ACTS Research
RESEARCH PROJECT · ACTS

Learning Who Talks to Whom: Role-Aware Communication Topology Scheduling for Efficient LLM Multi-Agent Collaboration

A role-aware policy selects communication modes and directed partners, then values the induced topology during centralized training.

Explore results
Role-aware scheduling · Directed communication · Centralized training
01 ABSTRACT

Communication is a design problem.

More agents do not automatically make a team better. The useful question is which agents should exchange information, in which direction, and when an additional exchange is worth its token cost.

ACTS learns a role-aware communication policy for LLM multi-agent collaboration. Local action values select a communication mode, while relational attention scores directed partners. These choices form a variable-cardinality graph for the next round.

During centralized training, a topological attention mixer encodes the current graph and assigns topology-dependent utility weights. Across nine benchmarks, ACTS reaches 75.01% macro accuracy with GPT-OSS-120B and 69.43% with DeepSeek-V4-Flash, improving over Agent Q-Mix by 3.18 and 6.19 percentage points.

Role-awareDirected topologySelective communication
FIGURE 1 ACTS treats communication topology as a learned decision: connect complementary roles when an exchange is useful, and stop when it is not.
02 APPROACH

Two decisions.
One adaptive graph.

ACTS separates partner selection from value estimation, then trains both through the graph they create.

Decentralized execution
Centralized training

ACTS at a glance

Open original
FIGURE 2 Local decisions choose communication modes and directed partners; the centralized mixer evaluates their joint value. Roles and node counts in the original illustration are schematic; experiments use three agents.
01

Choose who to contact

Role-conditioned recurrent states produce directional relevance scores. Communication modes and above-mean scores define the next graph Gt+1.

02

Value the graph in context

The mixer reads incoming and outgoing neighborhoods in Gt, assigning nonnegative weights to local utilities during centralized training.

Inside the dual-attention scheduler

Open original
FIGURE 3 Relational attention supports partner scoring; topological attention conditions value factorization on the realized graph. The current formulation uses Gt for the pre-decision graph and constructs Gt+1 from selected actions.
Illustrative topology sequenceThree-agent example
Communication decisions across roundsMethod illustration
03 EXPERIMENTAL RESULTS

Fewer exchanges.
Stronger answers.

Nine benchmarks, two frozen backbones, and the same evaluation protocol.

Download results
75.01%GPT-OSS-120B
macro accuracy
69.43%DeepSeek-V4-Flash
macro accuracy
+3.18 ppvs. Agent Q-Mix
GPT-OSS-120B
+6.19 ppvs. Agent Q-Mix
DeepSeek-V4-Flash
INTERACTIVE RESULTS

Inspect the reported numbers

Switch the backbone and task to compare accuracy with per-sample token use.

BACKBONE
+3.18percentage points

Higher macro-average accuracy than Agent Q-Mix
75.01% vs. 71.83% · GPT-OSS-120B

TABLE 1

Accuracy across nine benchmarks

Unweighted macro-average (%) · higher is better

01
ACTSAgent Q-MixOther evaluated methods

Accuracy meets token efficiency

Per-sample tokens vs. task accuracy

02
Upper left is better ↖
Hover or focus a point to inspect a method.
View exact values

Source: manuscript Tables 1 and 3. ACTS uses the fewest tokens among the evaluated multi-agent methods in these four settings; direct prompting uses fewer tokens overall.

Original paper figure and full benchmark tableView supporting evidence

Accuracy and token use

Open original
FIGURE 4 Accuracy–token comparisons across two tasks and two backbones. Several HLE points in the original graphic differ from Table 1; the interactive results use the numerical tables.
All methods. Every benchmark.View full results

Scroll horizontally to explore all nine benchmarks →

Accuracy (%). Bold values are best or tied best in each column. Avg. is the reported unweighted macro-average. HLE uses the first 250 questions.

EVALUATION SCOPE

Three role-conditioned agents and a final answer-aggregation node. Results are averaged over three independent random seeds across coding, reasoning and knowledge, multi-hop question answering, and competition mathematics.

04 ANALYSIS

Communicate when it helps.
Know when to stop.

The learned policy changes its communication pattern with the task.

SELECTIVE COMMUNICATION

Different tasks need
different conversations.

On HotpotQA, complementary evidence supports sustained interaction. On HLE, where the manuscript finds little accuracy uplift from collaboration, the learned policy deactivates agents more often.

4.2%Off actions · HotpotQA
21.3%Off actions · HLE

DeepSeek-V4-Flash · Section 4.3.2. “Off” skips the agent’s LLM call; “solo” keeps independent inference without communication.

Communicate when it helps

Open original
FIGURE 6 DeepSeek-V4-Flash selects off more often on HLE than on HotpotQA.
More analysesInformation gain, ablations, and action patterns

Where collaboration adds information

Open original
FIGURE 5 The benefit of collaboration varies by task and backbone; bars show topology-family means rather than ACTS alone.
COMPONENT ABLATIONS

What does each part contribute?

Removing a component changes both accuracy and cost

Macro-average accuracy and per-sample tokens over nine benchmarks.

Source: Table 2. The QMIX mixer replacement is a different comparison from the Agent Q-Mix baseline.

Task-dependent communication choices

Open original
FIGURE 7 The policy changes its communication profile across benchmarks.

Communication across rounds

Open original
FIGURE 8 Two original manuscript illustrations show how exchanges can change across rounds. These are schematic illustrations, not verified experimental trajectories.
ACTS · RESEARCH FIGURE