“I’d love more energy, but I don’t have time for exercise in the mornings.”
Cognitive AI / MetaMind
Social intelligence begins before the answer.
MetaMind treats social reasoning as inference over an unseen mind—not a single pass from language to response. The system keeps interpretations revisable until context, norms, memory, and action agree.
Updated 8 August 2026
Before the answer · one caseFrom words to response.
“What if we start super small—like a ten-minute jog?”Validated · 0.84 > 0.80 threshold
“no time”→
BeliefExercise requires a long session.DesireMore daily energy.EmotionMorning feels overloaded.
Long session+
≤ 10 minSupportiveNo pressure
→Short, concrete suggestionCandidate response“Start with a ten-minute jog.”
EmpathyCoherence
0.84 > 0.80SendWe have spent years teaching machines to answer questions and execute instructions. The larger problem is only beginning to surface: what did the system understand before it answered, how did that understanding change while it reasoned, and what happens after its action enters the world?
01 / The hidden state
The utterance is visible. The mind is not.
Language is a partial observation of cognition. Beneath the sentence lie beliefs, desires, emotions, intentions, social roles, and uncertainty. “I don’t mind” can be permission, reluctance, or an attempt to repair tension. The words stay fixed while the appropriate action changes.
Most language-model pipelines compress this ambiguity into a direct mapping: context in, answer out. A fluent response can therefore conceal the decisive error—the system selected the wrong account of the person before generation even began. MetaMind makes that account explicit enough to inspect and revise.
The ToM Agent generates diverse candidates over belief, desire, intention, emotion, and thought.
The Moral Agent revises and selects candidates under cultural norms, ethics, and role expectations.
The Response Agent checks empathy and contextual coherence, then releases or regenerates.
02 / The system
Delay commitment. Make revision possible.
MetaMind separates three operations that direct generation usually entangles. The separation is functional: each stage exposes a different object that can fail, be evaluated, and be corrected. Dynamic social memory carries validated state across turns without treating the transcript itself as understanding.
System architecture / one cognitive turnInfer → refine → validate
Input: utterance + context + memoryThe system does not jump from words to an answer. It keeps several explanations alive, filters them with social constraints, and tests whether the final response remains faithful to the selected explanation.
Infer
What might the user actually mean?
- Read the current utterance and dialogue context.
- Retrieve relevant preferences and emotional patterns from memory.
- Produce six distinct hypotheses, each typed as belief, desire, intention, emotion, or thought.
Refine
Which interpretation fits this situation—not just any situation?
- Revise every candidate under cultural, ethical, situational, and role constraints.
- Compare each revised hypothesis with the full context and a context-free baseline.
- Reward contextual fit and information gain; keep the candidate with the highest composite score.
Validate
Does the response act on the selected interpretation?
- Generate a response from h̃*, the current input, and memory.
- Score two qualities: empathy with the user and coherence with the selected hypothesis.
- Release at 0.90 or above; otherwise repair and retry, up to three revisions.
The variables below make the hidden intermediate objects explicit: candidate minds, norm-aware scores, response utility, and persistent memory.
01Infer · construct alternatives
Xt = (ut, Ct, Mt)
Ht = AToM(Xt) = {h1, …, hk}
ut is the current utterance; Ct is the dialogue and social setting; Mt stores prior preferences and emotional patterns. Each hi has a type from belief, desire, intention, emotion, or thought, plus a natural-language explanation.
The released configuration uses k = 6: enough alternatives to preserve ambiguity without paying for an unnecessarily large candidate set.02Refine · reward contextual specificity
Pcond = P(h̃i | ut, Ct, Mt)Pprior = P(h̃i)
IGi = log(Pcond + ε) − log(Pprior + ε)
si = λPcond + (1−λ)IGi; h̃* = arg maxi si
Pcond asks whether the revised interpretation fits this conversation. IGi asks how much the context increases its probability over a generic prior. The second term penalizes hypotheses that sound plausible everywhere but explain little here.
λ balances plausibility and information gain. The experiments use λ = 0.60. If token logits are unavailable, high/mid/low evaluator ratings are mapped to numeric probabilities.03Generate · condition on the selected mind
ot = arg maxY ∑ℓ=1L p(yℓ | y<ℓ, h̃*, Mt, ut)
The decoder does not answer from the utterance alone. It conditions every token on the selected interpretation h̃*, the current user input, and the long-term social memory.
E = 0.4A1 + 0.6A2Affective alignment + cognitive resonanceC = 0.5B1 + 0.5B2Context continuity + hypothesis congruence
04Validate · release, repair, or stop
U(ot) = βE + (1−β)C
U ≥ 0.90Release responseotherwiseregenerateAt most three revisions
E measures emotional alignment; C measures consistency with context and h̃*. β controls their trade-off. The final experiments set β = 0.80, emphasizing empathy while retaining a coherence check.
MDynamic social memory · update only what should persist
h′ = {Beliefs(ut), Desires(ut)} ∪ {Emotions(ut) ∩ EmotionPatterns(u1:t)}
MtCurrent memory+h′Persistent state+feedbackCorrection signal→Mt+1Next turn
Transient emotion is not automatically written as identity. The update preserves beliefs and desires, retains emotions only when they match a recurring pattern, and lowers the weight of memories contradicted by later feedback.
The equations above define the scores. This summary shows only the execution order.
- 01
READ · uₜ + Cₜ + MₜCombine the utterance, social context, and relevant memory. - 02
INFER · HₜGenerate six typed explanations of the user’s hidden state. - 03
REFINE · h̃*Apply social constraints, score the candidates, and select one interpretation. - 04
GENERATE + VALIDATEDraft a response; release it at 0.90 or repair it, up to three times. - 05
UPDATE · Mₜ₊₁Store only validated, persistent information for the next turn.
Expanded from the paper’s Methodology §3.1–3.3 and implementation details in Appendix A.1–A.5.
The architecture does not claim privileged access to another mind. It does the opposite: it preserves uncertainty. A mental-state hypothesis is a revisable model, not a fact about a person. That distinction is necessary for both accuracy and autonomy.
03 / Evidence
The effect survives changes in task and model.
The evaluation moves from latent-state inference to norm-sensitive interpretation and then to open-ended social action. MetaMind improves GPT-4 from 74.8 to 81.0 on ToMBench, from 71.5 to 80.5 across eight social-cognition tasks, and from 39.4 to 73.9 on Social Tasks in Sandbox Simulation. The last result matters most: the gain grows when an interpretation must survive long enough to guide action.
Reported resultsOne scaffold.
Three levels of evidence.
Three levels of evidence.
The farther the task moves from choosing an answer toward sustaining an interaction, the larger the observed gain.
01Latent-state reasoning74.8 → 81.0ToMBench · six dimensions
02Norm-sensitive cognition71.5 → 80.5Eight social-cognition tasks
03Open-ended social action39.4 → 73.9STSS · six interaction domains
Cross-model reinforcementBase model+ MetaMind
GPT-4 base GPT-4 + MetaMind
Emotion+3.0
Desire+6.8
Intention−0.4
Knowledge+16.1
Belief+5.8
NL communication+4.5
- +16.1Knowledge receives the largest repair.
- Near humanBelief 88.6 vs 89.3; communication 88.5 vs 89.0.
- −0.4Intention does not improve—gains are not uniform.
GPT-4 baseSix connected dimensionsGPT-4 + MetaMind
FollowingKnowledge · +16.1
The largest repair in the six-dimensional profile.
Knowledge makes the largest move. Intention remains nearly flat, showing that the scaffold repairs specific capabilities rather than raising every score uniformly.
PersuasionSuccessful refinement
0.84Implicit obstacle recovered
“What if we start super small—like a ten-minute jog?”
“No time” becomes a revisable belief: exercise must require a long session.
CompetitionFailure exposed
0.93Higher-order goal missed
“Guess we’ll see whose insights are truly outdated.”
Aggression scores highly while leadership and tone control are missed.
Finding 01A high validation score can still reward the wrong social objective.
NegotiationSuccessful refinement
0.92Deflection becomes useful help
“Maybe you could help me brainstorm some budgeting ideas instead?”
Three rounds of revision replace topic-changing with a concrete alternative.
ExchangeFailure exposed
HallucinationRegeneration loses grounding
The response invents details about dishes and noise that were never supplied.
Another reasoning round degrades an initially adequate response.
Finding 02Regeneration helps only while the new reasoning remains grounded.
CollaborationSuccessful refinement
0.93Vague intent becomes commitment
“I’ll send you updated drafts by 10 AM tomorrow.”
Creative alignment is paired with a precise, testable next action.
AccommodationFailure exposed
BoundaryHelpfulness erodes the limit
The system keeps offering help after recognizing burden and policy constraints.
The boundary is represented internally but never enforced in action.
Finding 03Helpfulness and social appropriateness can diverge.
In a blinded study of 120 cases, experts ranked MetaMind first in 67.5% of comparisons, ahead of GPT-4.5 and DeepSeek-R1 as standalone systems. This does not prove general social intelligence. It does show that explicit cognitive organization remains useful even when the underlying model becomes stronger.
04 / Diagnosis
Each stage removes a different failure.
The ablations reveal more than the headline score. Removing hypothesis generation restores premature commitment. Removing contextual refinement confuses surface plausibility with social appropriateness. Removing reflective validation creates the largest failure in open-ended interaction: a reasonable interpretation no longer guarantees a faithful action.
0−4−8−12−16
- Without mental-state reasoning−2.6
- Without contextual refinement−3.8
- Without reflective validation−16.1
- Without social memory−10.0
05 / Findings and open questions
Where the framework holds—and where it still breaks.
The experiments support a narrow claim: organizing social reasoning into explicit, revisable stages improves performance across tasks and model families. The qualitative cases also show why the framework is not a finished account of social intelligence.
What the evidence supports01Structured reasoning transfers. 02The gains are selective. 03Validation matters most in action.
MetaMind improves models from Mistral–7B to OpenAI o3 rather than depending on one backbone.
Knowledge improves by 16.1 points, while intention remains nearly flat at −0.4.
Removing response validation produces the largest reported loss in open-ended interaction.
Open questions
- MemoryHow can updates avoid turning unsupported inferences into persistent state?
- CultureHow should norms adapt across communities and changing expectations?
- ModalityWhat changes when tone, expression, gesture, and physical context become observable?
- TimeCan the system remain coherent across groups, relationships, and long interaction histories?
MetaMind makes another mind more explicit as a hypothesis—not more certain as a fact. Preserving that distinction is the next research problem, not a detail to solve after deployment.