Why AI Can Find a Fact but Lose Track of What Changed

Start with two scores: Ari has 10 points and Bea has 4. Ari gains 3. Bea’s score is set equal to Ari’s. Then Ari loses 2.

Ari now has 11 points. Bea has 13, not 11: her score copied Ari’s value at that moment, rather than staying linked to it. This is an illustrative example, not a test we ran on an AI. It exposes a distinction that matters when language models read long documents: finding a statement is not the same as following the changes that make it outdated.

Ai2’s new StateBench benchmark tests that distinction. In a preprint submitted November 2, researchers Jackson Petty, William Merrill and Ashish Sabharwal separate several kinds of bookkeeping that can look deceptively similar in a long prompt. Their experiments find that harder chains of updates remain a problem for the models they tested, even when simpler information retrieval works well.

The result is not a failure rate for everyday AI assistants. It is a clearer way to ask what a long-context model can actually do with the history it receives.

The Latest Mention Is Not Always the Answer

A fixed fact needs only a lookup. If a locker’s code never changes, finding the line that gives the code is enough.

An overwritten value is a little different. If a thermostat is first set to 68 degrees and later to 71, the last setting settles the question. Earlier settings can be ignored.

Accumulating changes remove that shortcut. A score that increases or decreases has to be carried through the updates. And when one score copies another, as in the example above, the correct answer can depend on a chain of earlier changes elsewhere in the document.

StateBench turns those distinctions into four controlled task families. It derives easier versions from a common sequence while keeping the number of variables, the available values and the positions of assignments aligned. The benchmark uses finite sets of values; it is not a test of arbitrary real-world arithmetic. The point is to change the kind of update rather than compare unrelated puzzles.

That helps separate two reasons a task might be hard: there is a lot to read, or many of the steps must be followed correctly.

The Harder Chains Expose a Different Weakness

The researchers tested four commercial frontier models by asking them to generate the final state after reading a sequence of changes. The models did well on the fixed-value and latest-value tasks. Performance deteriorated on the harder accumulating and dependent-update tasks as the sequences grew longer.

One missed update can matter much more in a chain than in a lookup. An old thermostat setting can be safely discarded once a new setting replaces it. An early change to Ari’s score cannot be discarded if Bea later copies the result. This is why checking whether a model can retrieve a fact from a huge document does not settle whether it can maintain the correct state through that document.

The paper also found a less obvious complication: presentation changed performance. Some smaller pretrained models handled a retrieval-style format well but struggled when the same underlying problem was presented as spreadsheet edits. A benchmark score can therefore reflect both the operation being tested and the model’s familiarity with its packaging.

That is another reason to resist treating a single score as a general measure of intelligence, a problem we explore in why AI benchmarks are not an IQ test.

Better Training Helped, but Did Not Erase Every Limit

To examine what the models could learn under more favorable conditions, the team ran a separate experiment: training small, 275-million-parameter models directly on StateBench tasks.

These models saw sequences of 8 to 256 assignments during training. Some recurrent and hybrid designs, which carry an internal memory forward as they process a sequence, handled the easier task families almost perfectly at lengths reaching 64,000 assignments. That is a result on those tasks, not evidence that a tiny specialized model beats a frontier chatbot at general work.

The hardest dependent-update family remained different. The more flexible Gated DeltaNet variant generalized well to roughly ten times its maximum training length, not all the way to the longest test sequences. The transformer-only variants fell short even on hardest-family sequences within their training range.

The contrast between direct training and ordinary pretraining matters. It shows why the observed failures cannot simply be labeled a universal architectural impossibility: what a model is trained to do, how the problem is written and how far the test extends beyond training all affect the result.

Keeping the History Is Only Half the Job

StateBench holds the set of variables fixed. Real conversations introduce new people, objects and obligations; real software may delegate bookkeeping to code or a database. The study does not establish how often these synthetic failure patterns appear in ordinary work, or how a complete assistant using such tools would perform.

It also asks a different question from persistent AI memory. Storing information between sessions determines what history an assistant can recover. Tracking state determines whether it can follow that history to the right current answer.

A bigger context window can put more of the record within reach. It cannot, by its size alone, establish that every consequential change in that record has been followed. For Ari and Bea, remembering all four instructions is useful. Keeping Bea’s score at 13 is the part that shows the updates were understood.