Further reading

Further reading

Where the mechanisms in this library come from. Nothing here is required to use it, but every design decision in tinyhivemind-hive traces to something on this page, and the Glossary links each term to the page that uses it.

The swarm mechanisms

Stigmergy. Coordination through traces left in a shared medium rather than through messages between agents. Named by Pierre-Paul Grassé in 1959 from termite mound construction.

• Stigmergy
• Swarm intelligence
• Self-organization

Used by: Trace grammar.

Pheromone trails and evaporation. A trail reinforced by use and weakened by time. Evaporation is the part that matters here: without it the first path found stays the only path found.

• Pheromone
• Ant colony optimization
• Exponential decay and half-life

Used by: the salience field in Hive episodes.

Quorum sensing. A decision taken when a threshold number of participants is detected locally, rather than by counting a majority globally. The term comes from bacteria; the version this library implements is the one honeybee swarms use to settle on a nest site.

• Quorum sensing
• Swarming (honey bee)
• Thomas Seeley, whose Honeybee Democracy is the book-length account of nest-site selection

Used by: the quorum fold in Hive episodes, and both of its bounds in Episode policy.

Cross-inhibition. Scouts advertising one site deliver stop signals to scouts advertising another, suppressing the advocate rather than devaluing the site. It is what lets a swarm break a tie between two good options, and the same asymmetry is why an !object in this library names a message rather than a topic.

• Swarming (honey bee)
• Lateral inhibition, the same shape in neural systems
• Seeley, Visscher, Schlegel, Hogan, Franks & Marshall, Stop signals provide cross inhibition in collective decision-making by honeybee swarms, Science 335:108–111 (2012), where the signal was shown to be targeted at scouts dancing for other sites.
• Pais, Hogan, Schlegel, Franks, Leonard & Marshall, A mechanism for value-sensitive decision-making, PLoS ONE 8(9):e73216 (2013). The equations: recruitment α(v)·x_u bilinear in the uncommitted pool, spontaneous abandonment ρ·x, and cross-inhibition β·x_A·x_B bilinear in the opposed populations. It is a leaky competing accumulator (Usher & McClelland, Psychological Review 108(3):550–592, 2001) with one extra constraint — a conserved population — and that constraint is what makes the swarm deadlock over two bad options and choose freely between two good ones.

The α(v) term is the one this library shipped without, and !refute is it: evidence bearing on an option changes what every member would independently conclude about the option, rather than silencing one advocate. The benchmark scored it and it lost; it is opt-in for that reason.

Used by: Hive episodes.

Response thresholds. Each individual has a private threshold for a task, acts when local stimulus exceeds it, and stays idle otherwise. Division of labour falls out of the threshold distribution without anybody assigning anything.

• Task allocation and partitioning of social insects
• Eusociality
• Bonabeau, Theraulaz & Deneubourg, Proc. R. Soc. B 263:1565–1569 (1996), for the fixed-threshold model s²/(s² + θ²); Theraulaz, Bonabeau & Deneubourg, Proc. R. Soc. B 265:327–332 (1998), for threshold reinforcement — doing a task lowers your threshold for it, so specialisation emerges from a homogeneous population and reverses when the specialist leaves.

Used by: the attention market in Hive episodes. The reinforcement half is not implemented: AgentThreshold.affinity is supplied by the host and never learned from the transcript.

Quorum responses, and why they filter cascades. Sumpter & Pratt, Quorum responses and consensus decision making, Phil. Trans. R. Soc. B 364:743–753 (2009). A quorum response is p = a + (m−a)·xᵏ/(Tᵏ + xᵏ) with k ≥ 2, which is sub-linear below the threshold — an active suppression of early social influence, built into the response curve. Marshall, Kurvers, Krause & Wolf, Quorums enable optimal pooling of independent judgements in biological systems, eLife 8:e40368 (2019), show simple majority voting is frequently sub-optimal and the optimal rule is a quorum at a threshold that is not 50%.

Speed against accuracy. Franks, Dornhaus, Fitzsimmons & Stevens, Proc. R. Soc. B 270:2457–2463 (2003): harassed ant colonies lower the quorum threshold, emigrate faster, and accept worse sites. The threshold is the dial, which is what Episode policy says about quorum.threshold.

Pandemonium. Oliver Selfridge's 1959 architecture, in which "demons" shout with a volume proportional to how well they match, and a decision demon takes the loudest. It is the response-threshold model arrived at from the artificial intelligence side rather than the entomology side, and taking the loudest one is exactly the argmax that enforces one message, one turn.

• Pandemonium architecture
• Arg max

Used by: Hive episodes.

How groups of people decide, and fail to

The failure modes this library designs against are not hypothetical, and none of them are specific to language models.

• Wisdom of the crowd and Condorcet's jury theorem, which is the formal reason an independent vote is such a strong control. The theorem assumes independence, and a shared transcript destroys it.
• Information cascade. The reason the opening round is blind. Without it, accuracy in the benchmark falls 24 points to the level of a single agent. See Benchmarks.
• Groupthink and the Asch conformity experiments. Convergence is a warning signal as much as a success signal, which is why the turn budget is deliberately small.
• Collective intelligence. Woolley et al., Evidence for a Collective Intelligence Factor in the Performance of Human Groups, Science 2010, found equality of conversational turn-taking to be one of the few robust predictors of group performance. That result is why the attention market carries a dominance penalty at all.
• Plurality voting, which is how the vote control arm decides.
• Hidden profiles. Stasser & Titus, JPSP 48:1467–1478 (1985). Groups discuss what everybody already knows and fail to surface uniquely-held information, so a group can systematically choose worse than its members' pooled information supports. This is what the checkout-503 live scenario is, and the failure recorded in the live run.
• Group size is not monotone. Kao & Couzin, Decision accuracy in complex environments is often maximized by small group sizes, Proc. R. Soc. B 281:20133305 (2014). When correlated cues dominate, accuracy peaks at a small finite group, because a large group reliably reproduces the population-mean opinion — which under a misleading shared cue is reliably wrong. This is the caution to read the "across desk sizes" table in Benchmarks with: those rooms have independent private evaluations, and real rooms share a brief.
• Diversity, carefully. Page's diversity prediction theorem — collective error equals average individual error minus prediction diversity — is an identity and is safe. The stronger "diversity trumps ability" claim is contested; see Thompson, Does diversity trump ability?, Notices of the AMS 61(9):1024–1030 (2014).
• The superorganism, honestly. Boomsma & Gawne, Superorganismality and caste differentiation as points of no return, Biological Reviews 93:28–54 (2018). A colony is not unitary — worker policing, worker reproduction, queen–worker conflict. Do not assume a room of agents has aligned objectives merely because one team built the system.

How shared context works when it works

The insect mechanisms say how a room converges. They are silent on what a trace should contain and how a reader knows it was understood. That is a different literature, and it is where most of what this library still lacks comes from. See Shared medium.

Transactive memory. Wegner, Transactive memory: a contemporary analysis of the group mind (1986). A group's memory is the directory — who knows what — rather than the contents, and it has three operations: directory updating, information allocation, and retrieval coordination. Lewis, J. Applied Psychology 88:587–604 (2003), validates the three factors: specialisation, credibility, coordination. A room that cannot route to the member holding the fact has a directory failure, not a reasoning failure.

Distributed cognition. Hutchins, Cognition in the Wild, MIT Press (1995), and How a cockpit remembers its speeds, Cognitive Science 19(3):265–288. Cognition is the propagation of representational state across media, and a system's cognitive properties follow from the physical properties of those media — which is the argument the Shared medium page makes about the transcript. The best shared memory turns a recall question into a perceptual one.

Grounding. Clark & Brennan, Grounding in communication (1991). Common ground is built by presentation and acceptance, to a criterion "sufficient for current purposes" rather than to identity of state. Their eight media constraints — reviewability, revisability, sequentiality, cotemporality and the rest — are a checklist for any transcript design, and Shared medium works through where this one lands on each.

Boundary objects and awareness. Star & Griesemer, Social Studies of Science 19(3):387–420 (1989): objects plastic enough to adapt locally and robust enough to keep a common identity — coordination without consensus. Gutwin & Greenberg, CSCW 11(3–4):411–446 (2002), on workspace awareness and its three channels: consequential communication, feedthrough (observing the effects of someone's work on a shared artifact — sematectonic stigmergy under another name), and expensive intentional telling.

Global workspace. Baars (1988); Mashour, Roelfsema, Changeux & Dehaene, Conscious processing and the global neuronal workspace hypothesis, Neuron 105(5):776–798 (2020). Parallel specialised processors compete for a capacity-limited workspace and the winner is broadcast to all of them, with ignition all-or-none, self-sustaining, and exclusive by lateral inhibition. That is bids → floor_holder → append, term for term. One message, one turn is not only a safety constraint; it is the architecture a brain uses.

Multi-agent language models

• Self-Consistency Improves Chain of Thought Reasoning in Language Models. Sample several answers independently and take the plurality. This is the matched-budget control in Benchmarks, and any multi-agent result reported without it is close to meaningless.
• Improving Factuality and Reasoning in Language Models through Multiagent Debate. The debate setup this library's episode is a bounded, auditable version of.
• Generative Agents: Interactive Simulacra of Human Behavior. The memory stream retrieval score, a weighted sum of recency, importance and relevance, is the shape the salience field borrows. Where its paper and its released code disagree on weights, this library follows the code.
• Why Do Multi-Agent LLM Systems Fail?. A taxonomy of observed failures. Step repetition and premature termination both appear in it, and both are protocol bugs rather than model bugs, which is why repetition_cap and the phase gate exist.

On communication topology: sparse topologies have been reported to match or beat fully connected ones in multi-agent debate at substantially lower cost. That finding, together with the conformity results above, is the empirical half of the argument in ADR 0002; the other half is that a type carrying two authorized speakers would make the charter's one-message-one-turn rule unenforceable.

Long context, and why the window is not the fix

Position bias. Lost in the Middle: How Language Models Use Long Contexts (Liu et al., TACL 2024). Accuracy is highest when the relevant document sits at the start or the end of the input and sags in the middle, and a longer window does not repair it. So enlarging a shared desk's window is not a free-with-upside change: a message that lands in the middle of a bigger window is less likely to be used than the same message at the edge of a smaller one.

Recursive language models. Recursive Language Models (Zhang, Kraska & Khattab, MIT CSAIL). Treat a long prompt as part of an external environment rather than as a prefix: hold the context in a REPL, and let the model write code to inspect it, chunk it, and recursively call itself over the parts that matter. It handles inputs two orders of magnitude past the window, and — the part that matters here — beats base models and long-context scaffolds even on prompts that would have fit, at comparable or lower cost. Querying a context can beat holding it when holding it was an option.

Recall is that move applied to a shared transcript rather than one prompt, and it stops well short of the paper: no recursion, no index, and a model call inside a fold would be a port, which the charter puts in the host. Nothing in the paper is about several agents sharing a log, and none of its numbers say anything about whether a room decides better — the same discipline the Benchmarks page applies to every other borrowed mechanism.

Used by: Recall, and BrevityPolicy, which states the window as a budget every message spends out of.

The machinery

Vocabulary this wiki uses without explaining it.

• Fold and pure function. Every decision in tinyhivemind-core and tinyhivemind-hive is one of these.
• Idempotence. Why an agent that catches up late folds to the same standing as one that watched live.
• Fixed-point arithmetic. Why every payload derives Eq and every episode replays exactly.
• Deadlock. What a below-majority quorum threshold produces, and why the threshold has a lower bound.
• Watermark. How continuous sharing tracks what an agent has already seen.
• Idempotency key. What the MentionTurnQueue port is keyed by.

In this repository

where	what
crates/tinyhivemind-hive/src/trace/	the stigmergic grammar
crates/tinyhivemind-hive/src/salience/	recency decay, importance, relevance
crates/tinyhivemind-hive/src/quorum/	standings, cross-inhibition, consensus
crates/tinyhivemind-hive/src/attention/	the bids and the argmax
crates/tinyhivemind-hive/src/episode/	the state machine and the visibility filter
crates/tinyhivemind-hive/examples/bench/	the benchmark harness
docs/adr/	architecture decision records
docs/specs/	behavior specifications
ROADMAP.md	the phase plan and the two defects it fixes