ZH’s prompt. Modeling/predicting human intent — building a metacognitive model — is arguably a core Meno goal. We work the algorithmic side (elicit high-quality human beliefs from behaviour + self-report); companies like Thinking Machines work the data side (capture high-granularity multimodal signal — speech, computer interaction — to read intent). Can we put both in the same camp for comparison? What are the metrics? Does it make sense to exhaust the data methods as a baseline to beat?

This page maps the landscape and argues the comparison is possible — but only after you split “intent” into distinct targets and fix a common held-out prediction task.

1. The crux: “human intent” is not one benchmarkable thing

Before any head-to-head, note that “intent” is at least five different targets, each with its own natural metric:

targetquestionnatural metricwho works it
Next-action / behaviourwhat will they do next?next-action accuracy, perplexity, calibrationtelemetry / digital-body-language, computer-use models
Goal / task recognitionwhat are they trying to accomplish?classification acc / F1, open-vocab matchmultimodal intent-recognition (AMCIU, IntentVLM)
Beliefwhat do they believe, and how will it update?Brier / corr(pred, stated), martingaleMeno (belief-elicitation); BigToM “beliefs-from-actions”
Preference / rewardwhat do they value?held-out choice prediction, reward recoveryMeno (LfL / ideal-preference); IRL
Real-time collaborative alignmentco-adapt to intent in the loopinteraction quality, turn-taking, proactivityThinking Machines interaction models

The single most important correction to the framing: Thinking Machines is mostly in the last row, not the intent-inference rows. Their Interaction Models optimise real-time interaction competence, not an offline “intent-accuracy” number (see §2). So a naive “who predicts intent better, TM or us?” is category-confused. The commensurable comparison is against the behaviour / goal / belief rows, where data-side predictors and Meno’s algorithmic models genuinely compete on the same held-out target (§4).

2. The data side — who, what, how well

Thinking Machines Lab — Interaction Models (May 2026). TML-Interaction-Small: a 276B-param MoE (12B active), trained from scratch so real-time full-duplex multimodal interaction is architectural, not bolted on. Continuous audio/video/text in 200 ms micro-turns; “encoder-free early fusion” (audio→dMel spectrograms, video→40×40 patches, shared transformer) + a parallel async background model for reasoning/tools. Target = the collaboration bottleneck: users who “can’t fully specify requirements upfront” and need continuous clarification.

  • Reported metrics (interaction, not intent): FD-bench 77.8; Audio-MultiChallenge 43.4%; turn-taking latency 0.40 s; internal TimeSpeak 64.7%, CueSpeak 81.7%, ProactiveVideoQA 33.5%, Charades 32.4% mIoU. Baselines = GPT-Realtime, Gemini Live.
  • Tellingly, they opened Interactivity Research Grants because the field lacks shared metrics for “good” real-time interaction — i.e. even the data-side leader admits the benchmark gap.

Multimodal intent-recognition research (the actual “read intent from signal” line).

  • IntentVLM — open-vocabulary intention recognition via forward-inverse modelling with video-language models.
  • AMCIU — adaptive multimodal complementary intent understanding, fusing gesture + speech + image via self-attention.
  • Multimodal intent classification + OOD detection — closed-set intent labels plus “is this a novel intent” detection.
  • Metric norm here: classification accuracy / F1 on labelled intent taxonomies — a much narrower target than belief.

“Digital body language” (product/UX). Hover, scroll, dwell, form-fills → funnel-stage / purchase intent (e.g. HCL). Correlational, coarse, but a real deployed baseline for behavioural intent.

Gaze / keystroke / EEG (implicit signal). Gaze-as-intent for device interaction; EEG-based implicit intent detection (“anticipation before action” in mixed reality); pupil modulation as a cognitive-control signal. High-granularity, low-level; strong on next-action, weak on why.

Personal-capture consolidation (worth noting for strategy). The consumer “capture everything you do → model you” play has largely been absorbed: Rewind → Limitless, acquired by Meta (Dec 2025), Pendant sales halted / Rewind sunset; Adept (computer-use) effectively went to Amazon (2024). The independent data-maximalist startup is a shrinking niche — the capability is being pulled inside the majors.

3. Existing benchmarks for intent / belief / ToM (and their metrics)

Almost all score endpoint QA accuracy, which is exactly the weakness Meno’s process-level (martingale/calibration) view exploits.

benchmarktargetmetricrelevance to Meno
SocialIQA (38k MCQ)intents + reactions to social scenesMC accuracybroad but shallow
BigToMbeliefs-from-percepts, actions-from-percepts, beliefs-from-actionsQA accuracy”beliefs-from-actions” ≈ Meno’s exact task
OpenToM1st/2nd-order beliefs + emotion, with personalitiesQA accuracynarrative, richer intents
OmniToMexplicit belief-structure modelling for all actorsbelief-state accclosest to “metacognitive model”
ToMATOverbalised mental states of role-playing LLMsQA accuracygenerative mental-state eval
InMindcapturing + applying individual human mindsetspersonalisation accindividual-level, like our sim-human

The user-centered ToM critique argues these endpoint benchmarks miss what matters for real assistance — which is precisely Meno’s opening.

4. Can data-side and algo-side share a benchmark? Yes — fix the task, not the method

The way to “put ourselves in the same camp” is to fix a common held-out prediction task on real human trajectories and let any method (data-maximal or algorithmic) compete:

  1. Behaviour prediction — predict the human’s next action/choice. Metric: accuracy / calibrated Brier / perplexity.
  2. Belief prediction — predict the human’s next stated probability (or eventual resolved outcome). Metric: Brier + corr(pred, stated) — Meno’s existing martingale machinery slots straight in.
  3. Preference prediction — predict held-out choices. Metric: choice accuracy / reward recovery.

Then compare methods on three axes, not one:

  • Predictive accuracy on the held-out target (the raw leaderboard — where data volume helps).
  • Sample / data budget — accuracy at matched data. This is where an algorithmic belief-model earns its keep: matching a telemetry-maximal model with far less signal.
  • Interpretability + intervention — does the model recover a belief/preference that is calibrated, causal, and steerable (predict behaviour under intervention, not just i.i.d.)? This is the active-identification angle — a correlational next-action predictor fails it; a genuine belief model shouldn’t.

On “exhaust the data methods as a baseline to beat”: yes, and it’s the right move — a data-maximal multimodal predictor sets the empirical ceiling for pure prediction. But framing our win as “beat their accuracy” concedes the wrong game. Our thesis should be: match the ceiling with less data, a calibrated/manipulable belief, and generalisation under intervention — accuracy parity + dominance on the sample-efficiency and interpretability axes. If we can’t even approach the ceiling, that’s a finding; if we match it with 10× less data and a steerable model, that’s the result.

5. Open questions

  • Which single held-out human-trajectory dataset could host this? (Candidates: our own reflection-game / gaokao logs; BigToM-style; a computer-use log corpus.) Do we build it or adapt one?
  • Is “intent” for our purposes best operationalised as belief (our comfort zone) or preference (LfL)? The metric follows that choice.
  • Interaction models (TM) are a different quadrant — is real-time co-construction of intent a fourth thing we should benchmark at all, or explicitly out of scope?
  • The intervention axis needs a protocol: how do you score “predicts behaviour under a counterfactual disclosure” on real humans without running the counterfactual? (Ties to active identification + the ethics of it.)

Sources