The Tower of Babel in artificial intelligence: who should speak, and when it pays to listen

Mohammad Salahshour

Nov 2, 2026arXiv:2611.03020v1

cs.LGcs.MA

Scorecard· 16/16

6.0/10 impact

Useful, honest conceptual separation of coordination vs cooperation with a solvable model and broad but stylized experiments; limited methodological novelty and far from real LLM-agent settings.

Abstract

Two machines can each recognize an object well and still fail together. We study classifiers that see different parts of one object and send short messages to a receiver that identifies its class. Coordination asks how to design messages together; cooperation asks whether the team outperforms one sender given the whole budget. On three digit datasets, messages designed together beat messages designed apart by 18.07 to 22.68 percentage points on average with two-bit messages, with gains in all 24 configurations. A fixed agreement, you report the row and I the column, recovers about half this gain, with only the receiver learned. Senders that draw the same distinctions waste a narrow vocabulary, and senders whose distinctions cross multiply it. A team has no more possible messages than one sender with the same budget, so a partner must bring evidence the first lacks. A solvable model shows individual statistics alone cannot determine whether a partnership helps. On MNIST digits, six shared bits remove 49.54 percent of the strongest tested solo's errors on average. Paired measurements correctly predicted most sensor teams would reduce average accuracy on new participants. Who should speak is a property of the pair, and must be measured in pairs.

AI Impact Assessments

(1 model)

Core Contribution

The paper studies two or more classifiers that each see part of an object and send bit-limited messages to a receiver that names the class. It separates two questions. Coordination asks how to design codes jointly. Cooperation asks whether a team beats a single sender given the whole bit budget.

The main findings:

Coordination gains. Joint code design beats independent design by 18–23 pp at two bits per sender on three digit datasets, positive in all 24 units. A combinatorial "coordination window" (K < C ≤ Kⁿ) explains where the gain peaks. A fixed row/column convention recovers about half the gain.

Cooperation is separate from coordination. A team has no capacity advantage over a solo with the same budget. A small XOR-style solvable model (Eq. 1, G* = ½[ρ − (1−2ε)]₊) shows identical individual statistics can coexist with zero or maximal team benefit. Empirically, teams beat the solo on digits (49.5% error removal on MNIST at six bits), but not on text halves, tabular data, or sensors, despite large coordination gains.

Prediction before deployment. Separate-sender calibration wrongly approved teams on text (predicted +1.69 pp, observed −6.18 pp) and dry beans. Paired calibration predicted the sign of sensor-team change on new people, though mostly losses.

Extras. A bound Dmarginal ≤ 2D*joint for Gaussian linear summaries, and a note that the optimal code depends on the loss.

Methodological Rigor

Strengths:

Careful data-role splits (representation, interface, selection, test).

Explicit competence thresholds.

A counting argument with clean ablations: independent-within-class views, random code pairs, and a fixed convention.

Forecasts recorded before testing.

Honest reporting of negative outcomes (the bird and paper pairs never met the competence bar and were not tested).

Bootstrap intervals for some key results.

Weaknesses:

Nothing was formally preregistered.

Several comparator choices were made post hoc (e.g., the multiple-features solo).

Groupings of outcomes were made after the results.

Spreads are descriptive, with few training splits (3 on MNIST).

Most sensor classifiers fail the competence bar, which confounds the interpretation of losses.

The sensor "prediction success" is weak: 5 of 6 are losses, and a rule that always predicts a loss would match five.

The "coordination" baseline (independent design as if each sender were the sole source) is a somewhat weak strawman. The large gains are partly by construction, since the counting argument makes them nearly inevitable. The headline 18–22 pp is also nearly a tautology of the code-crossing mechanism.

The setup is restricted to frozen classifiers, class-label messages, and tabular-receiver cells, so it is far from LLM agent communication despite the framing.

Potential Impact

The conceptual framing is useful: coordination gain does not imply cooperation gain; compare against the whole-budget solo; calibrate in pairs; account for unseen message combinations. It speaks to multi-agent LLM debates, where single-agent baselines under equal budgets are increasingly emphasized. It also touches distributed sensing and task-oriented communication.

The tools are simple enough to be adopted as an evaluation checklist. But the setting is stylized (digit datasets, class-label messages), the theory is modest (counting, a small XOR model, a factor-2 bound), and decentralized detection and coding-for-computing literature already contains much of the underlying intuition: joint quantizer design beats independent design, and sufficiency holds under conditional independence. The novelty is largely in the reframing and the empirical cataloguing of failure modes, not in a new method.

Timeliness

Highly timely: the multi-agent LLM literature increasingly finds that single agents match teams under equal budgets, and human–AI teams underperform their best member. The paper offers a clean mechanistic vocabulary for why. The link to actual LLM agents is indirect, since it uses frozen embeddings with linear readouts and small open models.

Strengths

Clear conceptual separation of two questions.

An exactly solvable model showing individual statistics are insufficient.

Pre-specified predictions, and candid reporting of failures and limitations.

Practical guidance: unseen-combination forecasts and sample-size needs for hypothesis tests (5.9% vs. 90.5% approval at 32 vs. 512 objects).

A broad range of datasets.

Limitations

Small, stylized tasks.

Limited baselines from the emergent-communication or learned-quantizer literature (end-to-end comparison only on MNIST).

No code is distributed beyond a "code-only package available from the author."

The long SI carries much of the evidence, so the main text is dense and hard to verify.

Some claims are weakly supported (sensor prediction).

Heavy reliance on AI assistants for analysis is disclosed, though the author verified the output.

The text and bird/paper negative results are inconclusive about cause, as the paper admits.

Writing is clear in places but dense, with many nested caveats and abbreviations.

Overall

A thoughtful, carefully hedged paper with a useful conceptual message and a decent empirical program. It is likely to be cited modestly as a framing reference rather than to become a foundation. Its novelty is moderate and its generalisability to LLM agent systems is unproven.

Rating:6/ 10

Significance 5.5Rigor 6.5Novelty 5.5Clarity 6

Generated Nov 3, 2026

Comparison History (0)

No comparisons yet.

Scorecard· 16/16

6.0/10 impact

Useful, honest conceptual separation of coordination vs cooperation with a solvable model and broad but stylized experiments; limited methodological novelty and far from real LLM-agent settings.