When AI Guesses What You'll Do, It's Usually Just Guessing
You've seen this trick before. Ask a language model what a character in a story will do next, and it gives you a confident answer. Ask it why, and you get a plausible-sounding justification stitched together after the fact, a story reverse-engineered to fit whatever it already picked. Most systems predicting human decisions pattern-match from situation to outcome and backfill the "mental" part for show.
Direct answering achieves F1 of 42.2. Six-sample self-consistency climbs to 56.8, which mostly buys the same guess with more confidence. Neither approach asks what the person in the scene knows, intends, or feels before acting.
Mental World Modeling proposes a different approach: treat beliefs, intentions, emotions, and norms as state variables that sit alongside physical facts from the first parsing step onward. The Mentis pipeline keeps a physical channel and a mental channel coupled through every step, from parsing the scene to simulating what happens next to scoring which action fits. Run in full, it reaches 66.8 F1 averaged across eight tested LLMs, a 24.6-point jump over direct answering and 10.0 points over self-consistency. The real question is where that mind-tracking earns its gain, and where the physical facts alone were doing the work all along.
The Two Channels Behind Every Decision
The physical channel tracks objects, characters, spatial relations, environment — where things are, who's in the room, what's touching what. Every world model already handles this.
The mental channel holds belief (what a character thinks is true), intention (what they're trying to do), emotion, disposition, norms, and constraints. Mentis represents this per-character: each person carries their own beliefs and goals. Actions come from both channels at once: a character's next move depends on where the object is and what they believe about where it is, and those two facts must update together when something happens.
The ablation numbers make the coupling argument concrete. Removing the mental channel entirely costs 12.1 F1 points. Removing the physical channel costs 16.5 points. But keeping both channels while decoupling them (letting physical and mental states update independently) costs only 6.4 points. If the mental layer's value came from merely existing, decoupling should hurt almost as much as removal. It doesn't. The extra damage from full removal versus decoupling is the interaction itself: belief updating in lockstep with the physical event.
GPT-5.4, Opus 4.8, and Fable 5 all showed the pattern, though not equally: decoupling cost them between 4.2 and 7.4 points each.
Keeping the physical and mental channels coupled preserves the 6.4 F1 points that independent simulation loses — the interaction between the channels, not their mere presence, carries the value.
Why Simulating What Happens Next Matters More Than Reading Minds
To find where a decision-prediction pipeline breaks, feed it the right answer at each stage, one at a time, and measure the score gain. This is the closest thing to an autopsy you can run while the pipeline is still live.
The oracle test hits four checkpoints where wrong guesses poison everything downstream: parsing the current state, generating the target's observation, simulating what happens after an action, and scoring the resulting branches. Swap in gold data at each one individually and measure the F1 gain.
The standout is gold successor states, the step where you simulate what the world looks like after each candidate action: +3.5 F1, the largest single-stage gain of the four and not close. The other three swaps each buy noticeably less.
Add the four gains together: +8.7 F1. Now hand the model gold data at all four stages simultaneously: +6.3 F1. The 2.4-point gap between "sum of isolated fixes" and "combined fix" is error propagation made visible. A wrong current-state parse and a wrong observation overlap: the same underlying mistake gets counted twice when you fix stages separately, and once when you fix them together.
Push the experiment further: give the full system gold intermediates at every stage and it reaches 97.0 F1. The human reference ceiling is 98.5 F1, leaving a 1.5 F1 gap that lands entirely on the final valuation step: the judgment call about which branch fits.
If you're building a system to predict what a person will do, the tempting place to invest is a richer taxonomy of their beliefs, wants, and social read on the room. The bigger payoff, on these numbers, is simulating accurately what the world looks like the moment after they act. That stage carries the most weight, and every other fix depends on it.
Fixing stages one at a time promises +8.7 F1; fixing them together delivers +6.3. The missing 2.4 points are the same underlying errors surfacing at more than one stage.
Some Decisions Are Easier to Predict Than Others
Break the results down by decision type and the spread is stark: interpersonal decisions hit 71.3 F1, object and resource decisions land at 66.2, spatial and perceptual decisions drop to 61.3, and risk or norm decisions come in at 59.6.
Seventy-eight percent of scenes involve two or more characters, so predicting the "right" action usually means predicting what one person believes about another. Interpersonal scenes are where that hidden layer of belief and intention does the most work, and exactly where the coupled physical-mental pipeline pulls furthest ahead. Risk and norm decisions hinge on judgment calls about acceptable behavior under uncertainty, thinner signal, harder valuation, lower ceiling. The gap between 71.3 and 59.6 maps how much a scene's outcome depends on what's inside someone's head versus what's sitting on the table.
The modality breakdown tells a related story, though the paper flags these as directional reads given the small subsets. Text records (320) score 68.1 F1, images (100) reach 66.2, and video (28 records) falls to 62.1. Text and stills give you a fixed frame to reason about. Video is different, scenes unfold over time, so the system has to constantly rebuild what a character could see at each moment as the world changes. Every frame adds another chance for that filtered view to drift from what the character would perceive. Text stays still. Video keeps moving.
Where the System Still Gets It Wrong
Before trusting the F1 numbers, ask whether the pipeline even produces a usable answer every time. Artifact presence for the current state checks out 94.6% of the time. Branch schema coverage, meaning all six candidate actions got simulated, holds at 89.7%. The options-only floor sits at 29.7–33.2% across all eight models, above the 16.7% of random guessing — and flat no matter how strong the guessing model is. The paper reads that flatness as a semantic-plausibility floor (some wrong options simply sound less like sane next actions), not a shortcut a model can exploit. The reasoning gap worth caring about is the distance from that ~31% floor to the full system's 66.8.
But 5.4% of the time, the current-state parse doesn't leave an artifact behind, meaning something downstream is scoring branches against an unverifiable state. Roughly one in ten records skips simulating a full set of options. Then there's perspective leakage at 8.0%: the pipeline builds a character's observation and slips in a fact that character has no way of knowing, a detail from another room, another character's private thought, something outside their line of sight. It's an epistemic access failure. When the whole system is built on tracking what each character believes based on what they can perceive, this problem shows up in one out of every twelve records.
Add the outright failure rate of 3.2% (records producing no schema-valid output) and you get a clearer read on the distance between a strong benchmark number and something you'd ship. A 66.8 F1 average sounds solid until you remember it's sitting on top of a pipeline that misfires structurally on roughly one in twenty runs and leaks information it has no business having on nearly one in twelve.
So What Should AI Model Next?
Two numbers matter most. Gold successor states buy +3.5 F1, the single largest oracle gain, which marks where to invest. And when all four stages get fixed together, 2.4 of the promised 8.7 points evaporate, because errors overlap across stages instead of adding up neatly. Fixes compound, and the earlier in the chain you make them, the more they're worth.
So that's where the engineering budget goes: get better at simulating consequences. A system that nails the belief model but botches the successor state is still going to hand you the wrong action, because the scoring stage judges branches that were never accurately projected in the first place.
Even granting the pipeline everything (gold parses, gold observations, gold successor states, gold branch scores) it tops out at 97.0 F1 against a 98.5 human reference. That remaining gap lives past all four fixable stages, in the final judgment call: which branch fits, once every fact feeding into it is already correct. Perfect information about the world still leaves the question of what a person would actually do with it. For anyone stacking interpersonal or social reasoning on top of an LLM world model, that's the part no amount of upstream cleanup buys back. The last mile is pure judgment about outcomes.
If you're building or evaluating a decision-prediction pipeline, run the same test: substitute ground truth one stage at a time and see which swap moves your score the most.
# Stage-wise oracle substitution, applied to your own pipeline:
# 1. Swap gold current-state parse in, model-generated elsewhere → measure ΔF1
# 2. Swap gold target observation in, model-generated elsewhere → measure ΔF1
# 3. Swap gold successor states in, model-generated elsewhere → measure ΔF1
# 4. Swap gold branch scores in, model-generated elsewhere → measure ΔF1
# Whichever ΔF1 is largest is where your pipeline is actually bottlenecked.
Top comments (0)