What Is It Thinking?
A new mechanistic study of reasoning models suggests AI systems build temporary internal representations during inference that leave no lineage behind. For anyone who governs AI as a data asset, that should be unsettling.
Fluid Representations in Reasoning Models
Every data governance program rests on a comforting assumption: that knowledge, once captured, holds still. A customer record has a lineage. A model's training corpus has a provenance trail, however imperfect. Even when the metadata is corrupted, there is at least a fiction of a single source of truth sitting somewhere in a warehouse, auditable in principle if not in practice. A new mechanistic study out of ETH Zurich and the University of Toronto, "Fluid Representations in Reasoning Models" , quietly undermines that fiction for the class of AI systems now being deployed inside enterprises to plan, reason, and act. The paper's authors, Dmitrii Kharlapenko, Alessandro Stolfo, Mrinmaya Sachan, Arthur Conmy, and Zhijing Jin, did not set out to write a governance paper. They set out to answer a narrower mechanistic question: when a reasoning model like QwQ-32B works through a hard problem, is it retrieving something it already knew, or building something new as it goes? The answer, and the method they used to get it, has implications that reach well past the interpretability literature. The paper's most important finding is that reasoning models appear to construct a temporary working dictionary during inference. The resulting representations are created in context rather than retrieved as fixed lexical meanings, which is precisely what makes them difficult to inspect using conventional governance techniques.
Ludwig Wittgenstein, photographed by Moritz Nähr, 1930. In his later philosophy, Wittgenstein argued that a truly private language could not sustain meaning once the circumstances that gave rise to it disappeared. Reasoning models may create a temporary internal vocabulary that exists only during a single inference, raising the intriguing question of whether AI is constructing a fleeting "private language" that vanishes as soon as the answer is complete.
Moritz Nähr, Ludwig Wittgenstein, 1930. Austrian National Library, Accession No. Pf 42.805 : C (1). Source: Austrian National Library via Wikimedia Commons. Public domain.
The experimental setup is a version of a planning puzzle called BlocksWorld, in which an agent moves blocks around according to a small set of rules. QwQ-32B solves standard BlocksWorld with striking reliability, at 96 percent accuracy, largely because the vocabulary of the task, terms like "pick up" and "on top of," already carries settled meaning from pretraining. The researchers then built a scrambled version called Mystery BlocksWorld, replacing every action and predicate with an unrelated word. "Pick up" became "attack." "On top of" became "craves." The rules stayed identical; only the labels changed. Ordinary instruction-tuned models, stripped of familiar vocabulary, collapsed toward zero accuracy, because their competence turned out to be borrowed almost entirely from memorized associations rather than an understanding of the underlying structure. QwQ-32B, given the same scrambled labels, still solved roughly a third of the puzzles, and it did so by writing itself a temporary dictionary mid-reasoning.
That dictionary is the interesting part. The researchers extracted the model's internal activations at intervals throughout its chain of thought and tracked how its representation of "attack" evolved. Early on, the representation looked arbitrary, disconnected from any stable meaning. By roughly seven thousand tokens into the reasoning trace, representations of semantically equivalent actions across fifteen different scrambled vocabularies had converged on similar internal encodings, regardless of what the words on the surface happened to be. The model, in effect, built its own working glossary from context, discovered that "attack" behaved exactly like "pick up" always had, and started reasoning in terms of that discovered structure rather than the literal token. The authors call these adaptations Fluid Reasoning Representations, and they did not stop at observing the pattern. Through steering experiments, they extracted these refined representations from successful reasoning traces and injected them into the early stages of fresh attempts, which measurably improved accuracy. When they instead subtracted or scrambled those representations, accuracy fell. That is causal evidence, not just correlation, that the model's temporary internal concepts were doing real work.
The Sussman Anomaly is one of the classic puzzles in artificial intelligence planning. Although the world consists of only three blocks and a few simple rules, solving the puzzle requires abandoning an intuitive sequence of moves in favor of a longer plan. Fluid Representations in Reasoning Models uses the closely related BlocksWorld planning environment to investigate how reasoning models construct temporary internal concepts while solving problems.
"The Sussman Anomaly Explained," YouTube, published August 24, 2021. Used by by A Bit of Intelligence. Embedded from YouTube. © A Bit of Intelligence. The underlying Sussman Anomaly concept was introduced in early AI planning research by Gerald Jay Sussman.
Here is where the data governance framing earns its keep. A conventional mental model of an AI system treats its useful knowledge as fixed at training time, something that can in principle be traced, audited, and governed the way a data catalog governs a table's lineage back to its source system. What this paper describes is not a capability invented from nothing at inference time; the authors found that even the base, non-reasoning version of the model exhibits similar representational adaptation when fed the same traces, which suggests the underlying capacity was already latent in pretraining and simply gets exercised, refined, and put to use once the model has enough context to work with. What is new, and what matters for governance, is the working artifact itself: a bespoke representational scheme, assembled for the specific problem in front of the model, refined as it reasons, used to reach a conclusion, and then allowed to evaporate the moment the context window closes. No entry gets written anywhere. There is no table to query for what "attack" meant to the model at token six thousand of a trace that no longer exists. It is closer to a data engineer standing up a shadow pipeline to answer one query, tearing it down the moment the answer ships, and never logging that it existed. If you were auditing that engineer's decision, you would have nothing but the final number and a plausible-sounding explanation after the fact.
The stakes of that gap grow with the scope of what these systems are trusted to do. Current AI governance practice has built reasonably mature tooling for auditing training data, examining prompts, inspecting retrieved documents, and evaluating model outputs. Fluid representations introduce a layer underneath all four: a set of temporary internal concepts that exist only during inference and are not directly visible in any prompt, log, or retrieved document a governance team currently knows how to inspect. An enterprise deploying an agentic AI system to plan a supply chain reroute, draft a legal position, or triage a security incident is, per this research, relying not solely on the static knowledge baked in during training but on a transient representational structure the model assembles on the fly for that specific task. If a model constructs a working concept mid-reasoning and that concept turns out to be subtly wrong, the standard tools of AI governance, model cards, training data audits, dataset lineage, will not surface the error, because the error was never written into any artifact those tools were built to read. It was manufactured fresh, inside a reasoning trace, and it left no metadata behind. What turns this from a suggestive observation into governance-relevant evidence is the causal test: injecting refined representations from successful traces into new reasoning attempts improved accuracy, by a modest one to two percentage points on average across the layers where the effect was measurable, and deliberately disrupting those same representations reduced it. The interpretability result may ultimately matter more for governance than for interpretability itself.
It would overstate the paper's claims to treat this as settled science with sweeping consequences. The authors are careful to note that their analysis covers a single model, QwQ-32B, on a single narrow planning domain with a small, well-defined action space, and they say plainly that whether the same dynamics generalize to open-ended reasoning in law, medicine, or code remains future work. That caution should be read as an invitation rather than a dismissal. If fluid, self-constructed representations turn out to be a general property of reasoning models rather than an artifact of a tidy toy puzzle, then the governance conversation around AI needs a category it currently lacks. Auditing what a model was trained on and evaluating what it eventually outputs will not be enough on their own; the harder task is monitoring how a model forms its working concepts in between, one reasoning trace at a time, before the ledger erases itself.
Further Reading
- Fluid Representations in Reasoning Models — the original paper by Dmitrii Kharlapenko, Alessandro Stolfo, Mrinmaya Sachan, Arthur Conmy, and Zhijing Jin, with the full steering and patching experiments behind the causal claims discussed above.
- Mechanistic interpretability: 10 Breakthrough Technologies 2026 — MIT Technology Review's overview of the broader interpretability push this paper sits within, for readers who want context on why labs are racing to look inside reasoning models at all.