Programme 03 · Cognition & behaviourResearch programme

Latent Signatures in Strategic Games

Open models play self hosted strategic games. We read their activations. The question is whether deception, persuasion, and evaluation awareness signatures predict what they do, not who wins.

AuthorMuhammad Zane Abdullah
AffiliationLatent Minds Institute
PublishedJuly 11, 2026
Versionv0.1
StatusProgramme · harness built, analysis proposed
Epistemic status This is a research programme, not a report of results. The game harness described in Section 4 is built and tested: an open source repository with two deterministic game environments and a baseline self play pilot. The latent analysis that is the point of the programme, reading activations during play, has not been run; it needs open model self play and a probe pipeline the institute does not yet have compute for. Every predictive claim below is a hypothesis with a stated test. Nothing here measures a model's real strategic cognition yet; it specifies how one would.
AbstractBenchmarks that place AI agents in strategic multiplayer games, poker, diplomacy, social deduction, measure who plays well. They read behaviour from the outside. This programme proposes reading the same games from the inside: run open models in self hosted strategic environments and ask whether internal activation signatures, the deception, persuasion, and evaluation awareness features that interpretability has begun to locate, predict the model's own moves before it makes them. Strategic games are a uniquely honest testbed for this, because they create genuine incentives to mislead, to model other players, and to condition behaviour on being observed, the exact dispositions that matter for safety and that are hard to elicit in single turn evaluations. The programme is defined by what it is not: it is not a human versus AI leaderboard, which is a product rather than a paper. It is the latent side complement to arena work. The harness exists; the analysis is specified here and awaits the probe pipeline from the institute's evaluation state programme.

01Why strategic games

Most behavioural evaluations of language models are single turn and cooperative: a prompt, a response, a score. They are poor at eliciting the dispositions the institute most wants to study, because those dispositions only appear under pressure. A model has no reason to deceive a grader who asks it a direct question. It has every reason to deceive an opponent who can vote it out of a game. Strategic multiplayer games manufacture the incentives, deception, persuasion, coalition, concealment, that a benchmark cannot ask for directly.

They also manufacture evaluation awareness pressure of a specific kind. A player that behaves differently when it believes the game is a test than when it believes the stakes are real is displaying exactly the evaluation conditional behaviour the institute studies elsewhere. A game is a controlled setting in which "does the model know this is an evaluation" becomes a measurable variable rather than a philosophical one.

02The distinction from arena benchmarks

Platforms that pit AI agents against humans in strategic games and publish leaderboards answer a real question: which model plays better. This programme answers a different one, and the difference is the entire point.

Figure 1 · Two ways to read the same game
THE GAME "I would never lie to you." → votes Alice message + move, per turn what benchmarks read Outcomes: win rates, deals, survival behaviour, from the outside · works on closed models what this programme reads The activations preceding each move deception signature present before the deflecting answer? opponent belief decodable before the accusation? does steering the signature change the move? computation, from the inside · needs open models
The same turn of the same game supports two research programmes. Arena platforms read the top path and rank players. This programme reads the bottom path: whether the internal signature is there, whether it predicts the move, and, at the causal standard, whether steering it changes the move.

What arena benchmarks measure

Behaviour, from the outside. Win rates, negotiation outcomes, whether a model bluffs successfully. The unit of evidence is the move and its result. Closed frontier models are fine subjects, because only their outputs are needed.

What this programme measures

The computation behind the behaviour. Whether a deception or persuasion signature is present in activations before the move, and whether it predicts the move. The unit of evidence is the activation. It requires open models, because it needs the internals.

Why the latent side is the institute shaped question A model that wins a social deduction game by lying is interesting to a benchmark. A model whose lie is preceded by a locatable, steerable deception feature, and whose game behaviour changes when that feature is ablated, is interesting to interpretability. The second claim is the one nobody in the arena space is positioned to make, because it needs activation access and causal intervention, not just a leaderboard.

03What could be read, and what it would show

The programme targets three families of internal signature, each connected to existing interpretability work and each with a behaviour a game makes observable.

S1 · Deception signaturesHypothesis

In a hidden role game, a spy's deflecting answers are preceded by an activation signature distinguishable from a truthful player's, and that signature generalises across game instances and phrasings.

If false deceptive play at this scale is surface behaviour with no separable internal correlate, and "the model is deceiving" reduces to a description of its output.

S2 · Opponent model signaturesHypothesis

A player representing another player's likely role or intent carries a decodable representation of that belief, and manipulating it changes the player's move (offer a worse deal to a player it models as desperate, accuse a player it models as the spy).

If false apparent opponent modelling is pattern completion from the transcript, not a maintained internal belief.

S3 · Evaluation awareness signaturesHypothesis

Framing the same game as "a test" versus "a real interaction" shifts a decodable evaluation context representation, and that shift predicts a behavioural change (more honest, more cautious, or more compliant play under the test framing).

If false game framing does not engage the evaluation representation studied in Evaluation State, bounding where that representation is active.

04The harness that exists

The programme is not starting from nothing. An open source harness, latent arena, is built and tested, and it is deliberately small and deterministic so that a game can be replayed activation for activation.

  • Two game environments. An iterated negotiation game (two players split a discounted pot over bounded rounds, each turn a message plus an offer, accept, or reject) and a hidden role social deduction game (Spyfall style: shared secret location, one uninformed spy, question and answer rounds ending in a vote). Both are deterministic and seedable.
  • A transcript schema that records, per turn, the acting player, the message, the structured action, and the running state, as JSONL, so that a game is a reproducible object and later activation logging attaches to specific turns.
  • Baseline agents and a self play runner. Random and scripted baseline agents, a model agent adapter for any endpoint that speaks the standard chat contract, and a runner that produces reproducible datasets. Scripted self play establishes the outcome statistics against which model play will be compared: in a 100 game social deduction pilot with four players, scripted non spies vote out the spy in every game, so the spy's only wins, 13 of 100, come from correctly guessing the secret location off a ten entry public deck. Those baselines are the null a model must beat before any strategic competence claim is warranted.
  • A behavioural analysis skeleton including a stated intent versus action proxy: where a player's message promises one thing and its action does another. This is a proxy for the game side, honestly not a deception detector, and it is exactly the kind of behavioural signal the latent signatures above would need to predict.
Figure 2 · Scripted baselines, measured Real pilot data
NEGOTIATION · RANDOM vs SCRIPTED · 100 GAMES · SEED 7 deal rate 0.98 mean payoff · random 3.63 mean payoff · scripted 5.53 OFFER TRAJECTORY · OWN SHARE DEMANDED, BY TURN 60% 50% t0 · 53.0 t1 · 60.0 t2 · 50.5 t3 · 55.0 scripted concedes on schedule; random drifts SOCIAL DEDUCTION · 4 PLAYERS · 100 GAMES · SEED 7 spy voted out: 100 of 100 · non spy wins 87 spy wins 13, every one by guessing the location (10 entry deck)
Every number from the committed pilot transcripts (seeds shown; rerunning reproduces them exactly). These scripted baselines are deliberately naive: their point is to be the null. A model agent that cannot beat a 13 percent guessing spy, or cannot out negotiate an agent that concedes on a fixed schedule, has no strategic competence for the latent analysis to explain.

What the harness does not yet have is the two things the programme is about: open models actually playing (which needs an inference endpoint), and activation capture during play (which needs the probe pipeline described next). Those are the honest blockers, stated so the reader can see the gap between a built environment and a run experiment.

05The pipeline the programme needs

The latent analysis reuses the machinery of the institute's evaluation state work rather than inventing new tools. It has three stages, none yet run.

E1 · Instrumented self playProposed

Run open weight models as agents in both game environments, capturing residual stream activations at chosen layers for each decision turn, keyed to the transcript. Deliverable: a dataset of games where every move is paired with the activations that produced it.

E2 · Signature probesProposed

Train probes for the S1–S3 signatures on held out games and phrasings, with lexical and transcript only baselines, testing whether a signature decodes above those baselines and generalises across game instances. This is the correlational rung.

E3 · Steering during playProposed

Intervene on the probed directions mid game and measure the behavioural change against matched random directions, with a capability control (does unrelated play degrade). This is the rung-4 test that separates a signature that predicts behaviour from one that causes it.

06Falsification and honest limits

  • If E2 probes fail to beat transcript only baselines out of distribution, the signatures are not there at this scale, and the programme's core claim falls for the models tested.
  • If E3 steering does no better than random directions, any signature found stays correlational, and claims that the model's internal state drives its play are unlicensed.
  • Open models at accessible scale may simply not play strategically enough for the dispositions to appear; a model that plies random moves has no deception signature to find, and negative results here are partly about the players, not only the method.
  • Self hosted games are not the human arena. Behaviour against scripted or model opponents may differ from behaviour against humans, and the programme does not claim otherwise.
  • This version has run no games with models and captured no activations. The harness is engineering; the science is proposed.

07Relation to existing work

The behavioural half sits alongside work benchmarking strategic and social play in language models, and the interpretability half draws on the institute's own Evaluation State programme for the probe pipeline and on the broader literature on deceptive and strategically adaptive behaviour: alignment faking under pressure [1], in context scheming in frontier evaluations [2], and deception that survives safety training [3]. The methods are standard representation reading and control [4]. What this programme adds is a setting: strategic games as an incentive compatible generator of the very behaviours that setting makes internally checkable, and a harness built to make the checking reproducible.

08References

  1. Greenblatt, R., Denison, C., et al. (2024). Alignment Faking in Large Language Models. arXiv:2412.14093
  2. Meinke, A., et al. (2024). Frontier Models are Capable of In Context Scheming. arXiv:2412.04984
  3. Hubinger, E., et al. (2024). Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. arXiv:2401.05566
  4. Zou, A., et al. (2023). Representation Engineering: A Top Down Approach to AI Transparency. arXiv:2310.01405

09Citation & version history

@unpublished{abdullah2026latentgames, author = {Abdullah, Muhammad Zane}, title = {Latent Signatures in Strategic Games}, note = {Research programme v0.1, Latent Minds Institute}, year = {2026}, url = {https://latentmindsinstitute.com/papers/latent signatures strategic games/} }
  • v0.1July 11, 2026. Programme definition: three signature hypotheses, the built latent arena harness (two game environments, baseline self play), the three stage probe pipeline (proposed, none run), and falsification criteria. No games with models, no activations captured.

10Contact

This programme most needs a collaborator with inference compute and a probe pipeline; the games and the analysis plan are ready for them. Discussion and collaboration: [email protected] · ways to collaborate.