Do Neurons Dream of Primitive Operators? Wake-Sleep Compression Rediscovers Schank's Event Semantics

Peter Balogh

cs.LGMar 26, 2026

Do Neurons Dream of Primitive Operators? Wake-Sleep Compression Rediscovers Schank's Event Semantics

Authors: Peter Balogh

Abstract

We show that they do. Roger Schank's conceptual dependency theory proposed that all human events decompose into primitive operations -- ATRANS (transfer of possession), PTRANS (physical movement), MTRANS (information transfer), and others -- hand-coded from linguistic intuition. We ask: can the same primitives be discovered automatically through compression pressure alone? We adapt DreamCoder's wake-sleep library learning to event state transformations. Given events as before/after world-state pairs, the system searches for operator compositions explaining each event (wake), then extracts recurring patterns as library entries under Minimum Description Length (sleep). Starting from four generic primitives, it discovers operators mapping to Schank's core: MOVE_PROP_has = ATRANS, CHANGE_location = PTRANS, SET_knows = MTRANS, SET_consumed = INGEST, plus compound operators (e.g., "mail" = ATRANS composed with PTRANS) and novel emotional-state operators absent from Schank's taxonomy. We validate on synthetic events, ATOMIC (Sap et al., 2019), and GLUCOSE (Mostafazadeh et al., 2020). On synthetic data, the discovered library achieves MDL within 4% of Schank's hand-coded primitives at 100% coverage (vs. Schank's 81%). On ATOMIC, Schank covers only 10%; on GLUCOSE, 31%. The discovered library covers 100% of both, dominated by mental/emotional operators -- CHANGE_wants (20%), CHANGE_feels (18%), CHANGE_is (18%) -- none in Schank's original taxonomy. Libraries discovered from one corpus transfer to the other with under 1 bit/event degradation despite different annotation schemes and domains, suggesting the operators are information-theoretically determined structure, not dataset artifacts.

Figures & tables

Figure 1: Analogy between BPE tokenization and operator discovery. BPE merges frequent character pairs into new tokens; our sleep phase specializes operators by relation and composes frequent operator pairs—the same compression principle applied to typed state transformations.

Category	Templates	Ground Truth
Transfer	give, sell, donate, steal	ATRANS
Movement	walk, drive, fly	PTRANS
Information	tell, teach, inform	MTRANS
Learning	learn, discover, realize	MBUILD
Emotion	become happy/sad/angry	state change
Consumption	eat, drink	INGEST

Table 1: Event templates and their ground-truth operator categories. Each template generates events with random entity assignments from pools of 24 people, 20 objects, 16 locations, and 16 information items.

ATOMIC Relation	State Change	Triple Pattern
xEffect, oEffect	State change	$(who,\text{state},\text{new\_state})$
xReact, oReact	Emotion change	$(who,\text{feels},\text{emotion})$
xWant, oWant	Goal change	$(who,\text{wants},\text{goal})$
xIntent	Intent	$(X,\text{intends},\text{goal})$
xNeed	Precondition	$(X,\text{needs},\text{item})$
xAttr	Attribute	$(X,\text{is},\text{property})$

Table 2: Mapping from ATOMIC relation types to state-change triples. Effects are further parsed into subtypes (location, possession, knowledge, emotion) via pattern matching.

Discovered	Type	Arity	Schank Equiv.	Example Verbs
MOVE_PROP_has	Specialized	3	ATRANS	give, buy, steal
CHANGE_location	Specialized	3	PTRANS	walk, drive, fly
SET_knows	Specialized	2	MTRANS	tell, teach, read
SET_consumed	Specialized ∗	2	INGEST	eat, drink
CHANGE_feels	Specialized	3	(novel)	cheer up, anger
UNSET_has	Specialized	2	ATRANS -	lose, discard

Table 3: Operators discovered via wake-sleep library learning. Specialized operators are relation-specific versions of seed operators. Compound operators are multi-step compositions. ∗ SET_consumed was pruned at medium scale but survives at large scale with sufficient frequency.

Figure 2: Example event decomposition. “John mailed Mary the book” is decomposed into ATRANS (possession transfer) composed with PTRANS (physical movement). The system discovers this decomposition from state diffs alone, without any linguistic knowledge.

Library	Ops	Events Explained	MDL	Ratio
Seed (generic)	4	500/500	3099	1.33
Discovered (pre-prune)	24	500/500	2408	1.03
Discovered (post-prune)	14	500/500	2334	1.00
Schank (hand-coded)	7	81/100 ∗	523 ∗	—

Table 4: MDL comparison across libraries. ∗ Schank’s library is evaluated on 100 events; it explains fewer events because our implementation does not cover all Schank primitives. The MDL ratio compares discovered vs. Schank on the same evaluation set, where the ratio is 1.04.

Dataset	Events	Time (s)	Operators	Iterations
Small	100	$<$ 1	6	3
Medium	500	165	14	5+prune
Large	2000	60 ∗	13	3

Table 5: Scaling behavior. ∗ Large dataset result is without pruning phase (specialization-only). The same core operators are discovered at all scales.

Figure 3: MDL learning curve across wake-sleep iterations. The core operators (ATRANS, PTRANS, MTRANS) emerge in iteration 1. Pruning removes 10 unused operators, further reducing MDL.

Ablation	Setting	Ops	bits/ev	Cov.	A/P/M
Baseline	4 seeds, $k$ =10, $B$ =30	14	27.4	100%	✓
Seed size	2 (SET, UNSET only)	12	26.4	100%	✓
Seed size	4 (+ MOVE, CHANGE)	14	27.4	100%	✓
Freq. thresh. $k$	$k$ =3	14	27.4	100%	✓
	$k$ =5	14	27.4	100%	✓
	$k$ =20	14	27.4	100%	✓

Table 6: Ablation studies on synthetic data (500 events except where

N

varies). “A/P/M” indicates whether ATRANS, PTRANS, and MTRANS analogues are discovered. The algorithm is remarkably stable: all three core primitives emerge in every setting tested, including datasets as small as 50 events.

Library	Ops	Total (bits)	bits/event	Coverage	$\log_{2}$ BF vs Seed
Seed (generic)	4	15,453	30.9	100%	—
Schank (hand-coded)	7	15,072	30.1	81%	+382
Discovered	14	13,807	27.6	100%	+1,647

Table 7: Bayesian MDL comparison (500 events). Bayes factors are computed as

2^{\Delta\text{bits}}

. The discovered library beats Schank by 1,265 bits (

\log_{2}BF=1265

, decisive evidence on Jeffreys’ scale).

Library	Ops	Total (bits)	bits/event	Coverage	$\log_{2}$ BF vs Seed
Seed (generic)	4	71,394	35.7	100%	—
Schank (hand-coded)	7	71,429	35.7	9.9%	$-35$
Discovered	15	70,014	35.0	100%	+1,380

Table 8: Bayesian MDL on ATOMIC data (2,000 events). Schank’s primitives cover only 9.9% of real commonsense events and are worse than even the generic seed library (

\log_{2}BF=-35

), because the unexplained-event penalty for 90% uncovered events overwhelms Schank’s specialized savings. The discovered library beats both alternatives decisively.

Discovered Operator	Usage	Dominant Relation	Schank Equivalent
CHANGE_wants	405 (20%)	wants	(not in Schank)
CHANGE_feels	357 (18%)	feels	(not in Schank)
CHANGE_is	355 (18%)	is (attribute)	(not in Schank)
CHANGE_state	271 (14%)	state	(generic)
CHANGE_needs	193 (10%)	needs	(not in Schank)
SET_has	160 (8%)	has	ATRANS

Table 9: Operators discovered from ATOMIC data, ranked by usage. The top five operators—comprising 80% of all usage—have no equivalent in Schank’s taxonomy.

Figure 4: Operator usage distribution in synthetic vs. ATOMIC data. Left : Synthetic data favors Schank’s physical-action primitives (MTRANS, PTRANS, ATRANS). Right : ATOMIC data is dominated by mental/emotional operators (red) not in Schank’s taxonomy, with physical-action operators (blue) accounting for

<

10% of usage.

Library	Train bits/ev	Test bits/ev	Gap	Test Cov.
Seed (generic)	34.9	29.7	$-$ 5.2	100%
Schank (hand-coded)	34.5	27.6	$-$ 5.9	10.2%
Discovered	34.2	29.6	$\mathbf{-4.9}$	100%

Table 10: Generalization results (1,600 train / 400 test). The negative gap indicates test events are slightly cheaper (smaller entity pool). The discovered library explains 100% of held-out events with the smallest generalization gap.

Operator	Synthetic Rank	ATOMIC Rank
SET_knows (MTRANS)	1st (24%)	9th (1%)
CHANGE_location (PTRANS)	2nd (24%)	10th ( $<$ 1%)
MOVE_PROP_has (ATRANS)	3rd (22%)	6th (8%)
CHANGE_wants	—	1st (20%)
CHANGE_feels	5th (5%)	2nd (19%)
CHANGE_is	—	3rd (17%)

Table 11: Operator rankings in synthetic vs. ATOMIC data. Schank’s primitives dominate synthetic data but are minor in naturalistic data, where mental/emotional operators dominate.

Library	Ops	bits/ev	Cov.	$\log_{2}$ BF vs Seed
Seed (generic)	4	33.6	100%	—
Schank (hand-coded)	7	38.9	31.3%	$-10{,}769$
Discovered	10	32.2	100%	+2,627

Table 12: Bayesian MDL on GLUCOSE data (2,000 events). Schank’s primitives cover only 31.3% of events and are decisively worse than the generic seed library (

\log_{2}BF=-10{,}769

). The discovered library achieves full coverage with strong MDL improvement.

Operator	Usage	Relation	Schank?
CHANGE_status	1,116 (55.8%)	status	generic
SET_has	405 (20.2%)	has	ATRANS
CHANGE_feels	228 (11.4%)	feels	no
CHANGE_location	153 (7.6%)	location	PTRANS
SET_knows	68 (3.4%)	knows	MTRANS
UNSET_has	30 (1.5%)	has	ATRANS -

Table 13: Operators discovered from GLUCOSE data, ranked by usage. Status changes dominate in narrative data, reflecting GLUCOSE’s story-grounded annotation scheme.

Library trained on	Tested on	bits/ev	Cov.
ATOMIC	ATOMIC	35.0	100%
GLUCOSE	GLUCOSE	32.2	100%
GLUCOSE $\to$	ATOMIC	35.7	100%
ATOMIC $\to$	GLUCOSE	33.1	100%

Table 14: Cross-corpus transfer. Libraries discovered from one dataset explain 100% of events in the other, with only marginal MDL increase (

<

1 bit/event). The operators generalize across corpora with different annotation schemes and domain distributions.

Computational Cognitive Modeling

Computational cognitive modeling uses formal models and simulations to explain how people perceive, learn, remember, reason, and choose.

See all 126 papers

ReasonOps: Operator Segmentation for LLM Reasoning Traces
May 28, 2026Daniel Lee, Owen Queen, James ZouLLM Reasoning Reasoning Trace Analysis
Stanford University
Why We Care About Understanding: Competence through Predictive Compression
Sep 4, 2026Matthieu Queloz, Pierre Beckmann
University of Bern, Department of Philosophy · École Polytechnique Fédérale de Lausanne (EPFL) · Idiap Research Institute
Do Language Models Track Entities Across State Changes?
May 28, 2026Zilu Tang, Qiao Zhao, Gabriel Franco +4LLM Interpretability State Tracking
Department of Computer Science, Boston University, Boston, USA · Department of Data Science, Monash University, Indonesia · Faculty of Computer Science, University of Vienna, Austria +1

May 28, 2026J/K move · Enter open · S save

ReasonOps: Operator Segmentation for LLM Reasoning Traces

Daniel Lee, Owen Queen, James Zou

Stanford University

LLM Reasoning Reasoning Trace Analysis CoT Evaluation CoT Reasoning Structured Reasoning Large Reasoning Models+4

Open paper AskPDF arXiv

Chain-of-thought traces from large reasoning models can span tens of thousands of tokens, yet we lack a vocabulary for describing their internal structure. Previous methods developed to analyze chain-of-thought traces are either too rigid or not expressive enough, failing to capture features across domains and models. To remedy this, we develop ReasonOps, an unsupervised, expressive method for annotating chain-of-thought traces, providing succinct universal operators. Using ReasonOps, we analyze 44,662 traces from 12 thinking LLMs spanning 6 families across 8 reasoning benchmarks and discover that they share a common compositional structure: 7 recurring reasoning operators -- discourse-level moves such as backtracking, inferring, and hypothesizing -- that emerge from unsupervised clustering of sentence-initial 3-token pivots. These operators appear across every model family and benchmark domain, confirmed by three independent LLM judges who classify held-out samples at 70 -76% accuracy. We analyze the structure of operators on easy vs. hard problems, revealing that reflective operators are more helpful on hard problems and harm performance on easy problems. Operator sequences are highly model-identifying: a classifier trained on operator distributions alone recovers the source model with macro-AUC, revealing that each model family has a distinctive reasoning fingerprint. Structural operator features predict within-problem answer correctness well above baselines. Classifiers built on these operators reach WP-AUC and on AIME specifically. ReasonOps further enables early quality estimation well before the trace completes: we predict at WP-AUC for only 50% of the trace. The ReasonOps pipeline is unsupervised and annotation-free, enabling deep insights into LLM reasoning traces as well as strong downstream results on model identification and correctness prediction.

Abstract

Computational Cognitive Modeling

Explore similar work

ReasonOps: Operator Segmentation for LLM Reasoning Traces

Why We Care About Understanding: Competence through Predictive Compression

Do Language Models Track Entities Across State Changes?