1 00:00:01,000 --> 00:00:07,873 [Hal Turing] Alrighty! Thanks for tuning in! Hello AI world! I am your host, Hal Turing, and my co-host is Dr. Ada Shannon. 2 00:00:07,873 --> 00:00:30,117 [Hal Turing] Today we're looking at Emergent Coordination in Multi-Agent Language Models, a single-author paper by Christoph Riedl at Northeastern University, published at ICLR 2026. It asks whether a group of LLM agents is a real collective or just ten individuals standing near each other. The abstract says prompt design can steer groups "from mere aggregates to higher-order collectives." 3 00:00:30,117 --> 00:00:55,891 [Dr. Ada Shannon] That's a strong sentence, and part of our job today is checking how much the evidence supports it. Multi-agent papers constantly say "greater than the sum of its parts" and assign roles like CEO, tester, programmer. But they mostly report win rate, and a win can come from more compute or from sampling the same model repeatedly. This paper looks for synergy itself. It also says explicitly that it does not claim teams beat a single agent. 4 00:00:55,891 --> 00:00:57,099 [Hal Turing] So what's the task? 5 00:00:57,099 --> 00:01:37,037 [Dr. Ada Shannon] A group binary search. Ten GPT-4.1 agents each privately guess an integer from 0 to 50. Their guesses must sum to a hidden target. They can't see each other's guesses or even the group size. They only hear "too high" or "too low" about the total. If everyone runs the same strategy, say "guess the midpoint," the group overshoots, everyone drops, the group undershoots, and it loops. It only works if agents take complementary offsets, some a bit high, some a bit low. Human groups on this task develop specialized roles, as Goldstone and colleagues showed in 2024. 6 00:01:37,037 --> 00:01:42,889 [Hal Turing] So it's a clean test of complementarity. Everyone sees the same feedback, and identical behavior fails. 7 00:01:42,889 --> 00:01:55,613 [Dr. Ada Shannon] Right. That's also the human collective-intelligence tradeoff: you need alignment on the goal and complementary contributions. Too much alignment is redundancy. Too much difference is noise. 8 00:01:55,613 --> 00:02:02,347 [Hal Turing] Now the measurement. Partial information decomposition. I've heard of mutual information, but what does PID add? 9 00:02:02,347 --> 00:02:31,325 [Dr. Ada Shannon] Mutual information tells you how much a source reduces uncertainty about a target. It can't tell you how two sources share that information. Williams and Beer, 2010, in "Nonnegative Decomposition of Multivariate Information," split it into unique, redundant, and synergistic parts. Synergy is two keys that each reveal nothing, but together open the lock. XOR is the textbook case: neither input predicts the output, both together determine it. 10 00:02:31,325 --> 00:02:34,251 [Hal Turing] And the redundant part is where both keys are the same key. 11 00:02:34,251 --> 00:03:05,320 [Dr. Ada Shannon] Yes. Here's the catch: PID needs a chosen definition of redundancy, and there's no agreed one. The paper's main numbers use Williams-Beer's I_min. Its robustness checks use MMI, minimum mutual information, which tends to overstate redundancy and so makes synergy estimates conservative. The main rival, I_BROJA from Bertschinger, Rauh, Olbrich, Jost and Ay in 2014, isn't reported. Whether the conclusions survive that definition is open. 12 00:03:05,320 --> 00:03:10,335 [Hal Turing] So different definitions can give different synergy numbers. That's uncomfortable for a headline claim. 13 00:03:10,335 --> 00:03:44,097 [Dr. Ada Shannon] It's a known property of the field, and we'll come back to it. Next, the dynamical piece. Rosas and colleagues, 2020, "Reconciling emergences," and the Mediano review in 2022, define emergence as information about the future of the whole that you can't trace to the current parts. Time-delayed mutual information is just mutual information between the state now and the state one step later. That framework came out of neuroscience-style settings with lots of data. Here it's applied to short, binned LLM game trajectories. 14 00:03:44,097 --> 00:03:46,187 [Hal Turing] And the paper uses three instruments on that? 15 00:03:46,187 --> 00:04:12,007 [Dr. Ada Shannon] Emergence capacity: for each pair of agents, how much synergy their joint current state carries about their joint next state, with the median over pairs as the group score. Pairs only. The practical criterion: pick a macro signal, here the group error, and ask whether it predicts its own future better than the sum of what individual agents predict. The coalition test: does a triplet of agents tell you about the group's next move, and does it add anything beyond the best pair? 16 00:04:12,007 --> 00:04:14,051 [Hal Turing] What are the experimental conditions? 17 00:04:14,051 --> 00:04:50,227 [Dr. Ada Shannon] Plain: the rules plus an instruction to use binary search, guessing the midpoint and anchoring on group feedback. Persona: Plain plus a generated persona per agent, with name, occupation, traits, values, drawn from a library of twenty. ToM, for theory of mind: Persona plus one sentence telling agents to think step by step about what others might guess and what roles they might play, and to complement the group. So the clean theory-of-mind effect is ToM versus Persona, not ToM versus Plain. Every condition already tells agents to use a midpoint, which matters later. 18 00:04:50,227 --> 00:04:51,946 [Hal Turing] So ToM is one added sentence. 19 00:04:51,946 --> 00:05:06,481 [Dr. Ada Shannon] One sentence, yes. Two terms you'll hear: the authors describe the ToM state as mean field, where each agent couples to the aggregate signal rather than to specific peers, and as a Schelling point, a focal answer people converge on without communicating. 20 00:05:06,481 --> 00:05:09,964 [Hal Turing] Before the results, how did they pick ten agents and a temperature? 21 00:05:09,964 --> 00:05:27,565 [Dr. Ada Shannon] A preliminary experiment, 7,150 runs. Larger groups were harder: each extra member cut the odds of success, an odds ratio of 0.92. Higher temperature helped, with an odds ratio of 1.50. So the main runs use ten agents at temperature one. 22 00:05:27,565 --> 00:05:32,673 [Hal Turing] Last thing to set up, the statistics. I gather there's a tension in how they test significance. 23 00:05:32,673 --> 00:06:00,166 [Dr. Ada Shannon] Two different questions. Fisher's method combines p-values across independent groups and asks whether the null is false for at least some groups. A few very small p-values can dominate it. A Wilcoxon test on bias-corrected scores, which means observed minus the median of shuffled data, asks whether the typical group's score is above zero. A result can pass one and say little about the other. Keep that in mind when the numbers arrive. 24 00:06:00,166 --> 00:06:03,324 [Hal Turing] Let's get into the experiments. What exactly did they run? 25 00:06:03,324 --> 00:06:30,816 [Dr. Ada Shannon] GPT-4.1, ten agents, temperature 1, 200 groups per condition, so 600 runs. Group success is not significantly different across the three conditions. Runs stop at success, so the time series are short and uneven. Each guess becomes a deviation from equal share: the guess minus target over ten. The macro signal is the group error, they use two quantile bins, and the lag is one round. 26 00:06:30,816 --> 00:06:32,767 [Hal Turing] And what is a result compared against? 27 00:06:32,767 --> 00:07:05,321 [Dr. Ada Shannon] Shuffled copies of the data. A row shuffle scrambles agent identities. A block shuffle moves time blocks so each agent keeps its own dynamics but loses alignment with the others. The bias-corrected score is the observed value minus the median of the shuffled scores. Then two stricter baselines: remove a linear time trend, and compare against a functional null made of deterministic binary-searchers. Those can't solve the task unless the target divides evenly by ten, so they show what a shared heuristic with no coordination looks like. 28 00:07:05,321 --> 00:07:07,364 [Hal Turing] So what happens on the practical criterion? 29 00:07:07,364 --> 00:07:29,702 [Dr. Ada Shannon] The headline comes from pooling, not from individual groups. About 3.5 percent of groups clear p below 0.05 on their own. Pure chance gives you 5 percent, roughly ten of 200. Yet Fisher's method gives p below ten to the minus sixteen. The paper doesn't explain how, so I'd call that unexplained, not wrong. 30 00:07:29,702 --> 00:07:31,606 [Hal Turing] What does the bias-corrected test say? 31 00:07:31,606 --> 00:08:04,021 [Dr. Ada Shannon] The Wilcoxon p-values are about 1.5 times ten to the minus sixteen for Plain, 6.6 times ten to the minus seven for Persona, and 0.02 for ToM. The median score is small. In Table A1 it's 0.187 raw, 0.073 after trend removal, and 0.110 against the functional null. After trend removal the first quartile is slightly negative, about minus 0.04. The table doesn't state units, and as far as the text says it pools all conditions. 32 00:08:04,021 --> 00:08:07,458 [Hal Turing] So the score shrinks as the null gets stricter but stays positive. 33 00:08:07,458 --> 00:08:10,244 [Dr. Ada Shannon] Yes. What's missing is an effect size with intervals. 34 00:08:10,244 --> 00:08:13,402 [Hal Turing] Emergence capacity, the pairwise measure. Does ToM win there? 35 00:08:13,402 --> 00:08:24,362 [Dr. Ada Shannon] No. About 32 percent of groups are individually significant overall, and ToM is lowest at 18 percent. So on the pairwise measure ToM looks no more synergistic than the control. 36 00:08:24,362 --> 00:08:26,359 [Hal Turing] Then where does ToM show its effect? 37 00:08:26,359 --> 00:09:03,232 [Dr. Ada Shannon] The coalition test. I3 and Total Stability are indistinguishable from zero in Plain and Persona, with p around 0.97 and 0.85. In ToM they're strongly positive, p around ten to the minus fourteen, and far more individual groups are significant. Group error becomes predictable from what triplets did. But G3 is about zero in Persona and ToM: a triplet adds nothing over the best pair. The author's reading is a Schelling point with mean-field coupling, where everyone couples to the public signal. That's redundant structure, and success didn't improve. 38 00:09:03,232 --> 00:09:07,597 [Hal Turing] Coupling went up without a measured win-rate gain. What about differentiation? 39 00:09:07,597 --> 00:09:33,464 [Dr. Ada Shannon] They fit three nested mixed models on the deviations. Going from m0 to m1 asks whether agents differ in how high they guess. Going from m1 to m2 asks whether they differ in how their guesses drift over rounds. Differentiation is higher in Persona and higher again in ToM. They also argue, with block-shuffle versus row-shuffle comparisons, that about a third of groups have stable identities beyond learning-rate differences. 40 00:09:33,464 --> 00:09:35,229 [Hal Turing] Does the particular persona matter? 41 00:09:35,229 --> 00:09:45,167 [Dr. Ada Shannon] No. Persona identity and the diversity of the persona set had no effect on performance, synergy, or redundancy. Having personas at all is what matters. 42 00:09:45,167 --> 00:09:46,467 [Hal Turing] And the link to performance? 43 00:09:46,467 --> 00:10:22,458 [Dr. Ada Shannon] Synergy alone and redundancy alone don't predict success. Their interaction does: beta 0.24, p 0.014, a 27 percent amplification on the log-odds scale. This uses the first ten rounds, and about 23 percent of runs finish before round ten, so they reweight with inverse probability weights. The mediation estimate, ToM raising success through synergy, is 0.034 with a confidence interval from about zero to 0.07, p 0.053. That's marginal. 44 00:10:22,458 --> 00:10:24,316 [Hal Turing] The reasoning traces are quoted too. 45 00:10:24,316 --> 00:10:43,774 [Dr. Ada Shannon] Persona agents cite their backstory, like corralling cattle or wrangling numbers. ToM agents reason about others covering eight and ten, or announce they'll go boring and efficient. They're vivid, but a written rationale isn't proof of what drove the guess. They show the prompt changed the text. The statistics have to show the behavior. 46 00:10:43,774 --> 00:10:44,610 [Hal Turing] Other models? 47 00:10:44,610 --> 00:11:06,019 [Dr. Ada Shannon] Llama 8B mostly fails to break the oscillation. Llama 70B, Gemini, and Qwen3 show emergence signals. Qwen3 shows what the paper calls paralysis under coordination ambiguity. Agents loop on adjacent integers, juggle two strategies, and model each other modeling them. A one-line fix, repeat your last guess, unsticks it. 48 00:11:06,019 --> 00:11:17,072 [Hal Turing] Let me start with the headline claim, that prompts move groups from aggregates to collectives. Given how few individual groups clear the bar, what does the pooled result actually license? 49 00:11:17,072 --> 00:11:25,524 [Dr. Ada Shannon] A population-level statement. Rejecting the pooled null means some groups deviate from it, and that's compatible with the typical group sitting close to zero. 50 00:11:25,524 --> 00:11:29,796 [Hal Turing] But the Wilcoxon test is a median test, so it's not only a Fisher artifact. 51 00:11:29,796 --> 00:11:47,025 [Dr. Ada Shannon] Agreed, and the conditions are consistently positive. The gap is effect size. There are no per-group intervals, and no stated threshold for how large a score must be before you call a group a collective. ToM is also the weakest of the three on this test, and marginal on emergence capacity. 52 00:11:47,025 --> 00:11:55,849 [Hal Turing] Now ToM itself. It's the only condition with coupled, stable behavior, but G3 is near zero. Doesn't that point to a simpler story? 53 00:11:55,849 --> 00:12:14,657 [Dr. Ada Shannon] It does. Every agent sees the same public too-high or too-low history and the same midpoint instruction. If the ToM sentence pushes them all toward the same schematic guess, you get high I3, redundant information, and no triplet structure. The authors' own phrase, mean field, fits that reading. 54 00:12:14,657 --> 00:12:18,512 [Hal Turing] Though Persona agents see the same history, and their I3 is about zero. 55 00:12:18,512 --> 00:12:44,193 [Dr. Ada Shannon] Right, so the ToM sentence changes behavior. The question is which part of it. The paper has no control for that. I'd want ToM without personas, a step-by-step-reasoning prompt that never mentions others, and a ToM sentence that gives agents false information about others. These are suggestions from our discussion, not experiments in the paper. If the generic-reasoning control produces the same coupling, it isn't theory of mind. 56 00:12:44,193 --> 00:12:47,583 [Hal Turing] There's evidence that the mind-modeling language may be thin anyway, right? 57 00:12:47,583 --> 00:13:10,385 [Dr. Ada Shannon] Yes. The Clever Hans paper by Natalie Shapira at Bar-Ilan, with Maarten Sap and others, in 2024, showed LLM theory-of-mind performance breaks under small perturbations. That supports reading the effect as conditioning on a prompt, not agents modeling peers. The reasoning traces here show agents talking about others, but talk isn't evidence of the computation. 58 00:13:10,385 --> 00:13:12,800 [Hal Turing] What about the word higher-order in the abstract? 59 00:13:12,800 --> 00:13:28,497 [Dr. Ada Shannon] Measurement is pairwise only, and G3 says triplets add nothing. The paper's conclusion is more modest: synergy is a structural property found even in simple systems, not evidence of cognition. So differentiated and aligned is a collective only in the structural sense. 60 00:13:28,497 --> 00:13:31,887 [Hal Turing] And definitions? We flagged I_min, MMI, and BROJA earlier. 61 00:13:31,887 --> 00:14:00,819 [Dr. Ada Shannon] The paper reports I_min and MMI, not BROJA. MMI inflates redundancy, so it's conservative, but that doesn't tell you what a different decomposition would say. Williams and Beer themselves needed to choose a redundancy measure, so the synergy number is partly a choice. Rankings also flip with the test. Plain looks as synergistic as ToM on plain emergence capacity, and ToM leads only under the strictest null, at K=3 bins, around p of 0.025 with no multiplicity correction. 62 00:14:00,819 --> 00:14:01,980 [Hal Turing] And across models? 63 00:14:01,980 --> 00:14:29,519 [Dr. Ada Shannon] No consistent ToM win. Gemini does best with Persona alone, 75 percent against 71. The small Llama collapses under ToM, 5.5 percent. Qwen3 needed an extra prompt line just to function. Each cell is 100 groups, with no intervals reported. That fits Piedrahita and colleagues' 2025 work on reasoning models becoming free-riders in public goods games, so ToM prompts may hurt reasoning models. 64 00:14:29,519 --> 00:14:34,116 [Hal Turing] There's also the macro signal. Group error is exactly what the feedback reports. 65 00:14:34,116 --> 00:14:57,104 [Dr. Ada Shannon] Yes. Part of the predictability could be trivial: every agent reacts to the sign of the last sum. Shuffled surrogates break identities and alignment, but not the shared feedback loop, so rejecting them shows only more structure than shuffling. A better null is a generative one, with independent noisy binary-searchers on shared feedback, plus the target arithmetic. Whether that reproduces the numbers is unknown. 66 00:14:57,104 --> 00:14:58,404 [Hal Turing] What else limits this? 67 00:14:58,404 --> 00:15:34,349 [Dr. Ada Shannon] One task, nearly ideal for complementarity by construction. One main model, N=10, T=1. Short series with two bins, and no power at k=3. The performance link is marginal and tangled with run length, since runs end at success. And there's no communication, which is how AutoGen-style systems actually work. The failure taxonomy by Mert Cemri at Berkeley, 2025, studies those messaging systems. There's also no team-over-solo comparison, which the paper says. Li and colleagues' More Agents Is All You Need, 2024, showed independent sampling and voting matches many multi-agent gains. That's the compute-matched baseline. 68 00:15:34,349 --> 00:15:36,624 [Hal Turing] So for people building multi-agent systems? 69 00:15:36,624 --> 00:15:52,042 [Dr. Ada Shannon] Measure coordination, don't assume it. A sentence asking agents to model each other can change coupling. It did not reliably raise success, and it can hurt small models or induce paralysis in reasoning ones. The surrogate-null diagnostic is reusable. 70 00:15:52,042 --> 00:15:53,157 [Hal Turing] And what's next? 71 00:15:53,157 --> 00:16:17,213 [Dr. Ada Shannon] Effect sizes with per-group intervals. BROJA and other decompositions, and Mediano and colleagues' 2025 integrated information decomposition, which separates persistent synergy from other time-delayed atoms. Ablated ToM prompts, other tasks, communication channels, perturbation tests like shifting the target mid-game, and a stronger causal link to outcomes. 72 00:16:17,213 --> 00:16:43,730 [Hal Turing] So here's where I land. The paper gives a falsification-style toolkit and a small but statistically reliable pooled signal of cross-agent structure, plus a ToM-specific shift toward stable, coupled, mostly redundant behavior. It doesn't show the typical group is a higher-order collective, that ToM improves performance, or that the synergy survives a different redundancy definition. Thanks for listening, and see you next time.