Sleep That Remembers Its Shortcuts: Provenance-Tracked Consolidation for Language Models on Evolving Knowledge Benchmarks

When people rest after learning that A leads to B and B leads to C, their brains replay those memories and build a direct shortcut from A to C. The shortcut makes inference fast, but it is rigid. When the world later changes so that B now leads to C*, people who formed the shortcut keep reaching for the old C. Language models that are periodically updated on fresh text face the same problem. A model can learn that a company's chief executive changed and still answer a two-step question ("which university did the CEO of X attend?") with the old executive's school, because the composed answer was stored separately from the fact it depended on. Benchmarks that refresh their questions over time are the natural place to measure this failure. Today they mostly test isolated facts, and the consolidation methods that might fix it have only been tested on static image-classification splits.

We propose an offline "sleep" phase for a language model. It runs between successive snapshots of an evolving knowledge benchmark. During sleep the model replays new and changed facts, writes two-step compositions into its weights as direct associations, and records for each composition the facts it was derived from. That record is its provenance. When a later snapshot flags one of those facts as changed, the model invalidates and re-derives only the compositions that depended on it. The core claim is this: a model with dependency-targeted reconsolidation answers multi-step questions about changed knowledge as well as a model that regenerates all of its compositions from scratch every cycle, at a small fraction of the compute. It also clearly beats both plain continued training and inference-time retrieval on the "stale composition" failure. To test this, we build a multi-hop extension of an existing living benchmark whose questions are designed to expose stale shortcuts.

What the two literatures give us

The consolidation side supplies both the motivation and a warning. Two human studies supply the motivation. Memory reactivation during rest forms shortcuts in a cognitive map showed that cueing memories during awake rest turns A→B and B→C into a direct A→C shortcut (p = 0.013); participants' eyes skipped the intermediate item. Memory reactivation generates new, adaptive behaviours that reach beyond direct experience found that rest replay improves untrained multi-step inferences while leaving trained single-step associations unchanged.

The companion report, Memory reactivation during rest forms shortcuts in a cognitive map, supplies the warning that shapes our design. After a "value flip" that changes B→C to B→C*, people with consolidated shortcuts kept fixating on the obsolete outcome (p = 0.002). We read this as a prediction for machines: consolidating compositions buys speed and accuracy on stable knowledge, and pays for it when that knowledge drifts. Eran Eldar et al. describe offline replay as compiling flexible, model-based planning into fast cached policies. That framing also implies the cache must be invalidated when the underlying model changes.

On the machine-learning side, Wake-Sleep Consolidated Learning alternates fast "wake" learning, which fills a short-term buffer, with "sleep." Its sleep has a replay stage and a "dreaming" stage on an external, out-of-distribution image set. It reports gains of up to 12 points of final average accuracy over standard rehearsal on Split CIFAR-10. We borrow its wake/sleep structure but change the dreams. Instead of an external corpus, our dreams are compositions the model generates from its own buffered facts, each tagged with provenance.

When to Learn and When to Stop is a constraint. On Split MNIST, sleep replay alone peaked near 60% but swung wildly between phases, while gating on global validation accuracy gave a smoother curve near 45%. We therefore gate every sleep update. Our gate cannot assume a fixed validation set, because in our setting the questions rotate. REMIND Your Neural Network to Prevent Catastrophic Forgetting shows that compressed replay of stored internal activations scales to over a million instances. It assumes a frozen feature extractor, which we do not, but it suggests our replay buffer can be kept cheap.

The evolving-benchmark side supplies the testbed and the evaluation discipline. GrowOVER tracks monthly Wikipedia snapshots through 2023. It labels facts Unchanged, Changed, or New using sentence-similarity and entailment models, and builds question-answering sets with evidence passages. With Llama2-7B, its retrieval system RiLM scored 44.9 F1 against 43.4 for continued pretraining (RaLM-CP) (GrowOVER). That result motivates us: parametric updating currently loses to retrieval. The benchmark's limitation also motivates us: it evaluates isolated single-article changes, not how a change ripples through dependent facts.

LiveBench contributes the methodological rule we adopt, which is to grade programmatically rather than with an LLM judge. The Benchmark Lottery shows that rankings flip under small changes in which tasks are included. We therefore report per-category results and variance rather than one aggregate.

SocioDojo is the closest "lifelong agent" work. Its agent relies entirely on tools and scratchpads and never updates its weights. We leave it as a later application.

None of these papers puts a consolidation mechanism inside a living benchmark. None measures whether consolidated compositions go stale.

Research Program

The model is a 7B open-weights language model, matching GrowOVER's Llama2-7B setting so that published numbers stay comparable. Each GrowOVER monthly snapshot is one day.

Wake phase. The model answers that month's questions, with or without retrieval. New and changed passages go into a short-term buffer.

Sleep phase. Sleep has three stages.

  1. Replay. The model trains on buffered facts, mixed with a sampled stream of unchanged facts from a long-term store.
  2. Self-generated dreaming. The model links entities shared between buffered facts and long-term facts, generates candidate two-step compositions, and checks them against the source passages by string and entity matching. It then trains on the surviving compositions as direct question-answer associations. Each composition is stored with the identifiers of its supporting facts.
  3. Reconsolidation. For every fact the GrowOVER pipeline marks Changed, the model looks up all stored compositions whose provenance includes that fact. It re-derives them from the updated fact and trains on paired examples in which the new composed answer is correct and the old one is explicitly disfavored. This last ingredient matters because adding the new answer does not by itself remove the old association.

Gating. A sleep update is accepted only if accuracy does not drop on a rotating probe slice drawn from Unchanged knowledge and earlier compositions. This adapts the global gate of When to Learn and When to Stop to a validation set that itself refreshes.

Benchmark extension. We build GrowOVER-Hop. For each month we take pairs of GrowOVER facts that share a linked entity and form two-step questions with programmatically checkable entity answers. We then sort them into three groups:

  • stable: neither hop changed;
  • flip: exactly one hop changed between snapshots, the analog of the human value flip;
  • new: at least one hop is new knowledge.

Every flip item also has its single-hop probes. This lets us compute the key diagnostic, the stale composition rate: among flip items where the model answers the updated single hop correctly, the fraction where its two-step answer is still the old composed answer. This separates "did not learn the change" from "learned it but did not propagate it."

Baselines

The naive reference is the frozen base model with no updates, evaluated on every snapshot. It sets the floor for Changed and New knowledge and the ceiling for retention.

The strongest published competitor on single-hop evolving knowledge is RiLM, from GrowOVER. It is a training-free retrieval system with a certainty gate (about 75% accurate) that triggers re-retrieval. It reports 44.9 F1 on GrowOVER-QA with Llama2-7B. We will re-run it with the same base model, retriever, and document snapshots. Beating it on GrowOVER-QA is not required. Beating it on flip items in GrowOVER-Hop is required. On multi-hop flip questions there is no reported result on Lacuna for RiLM or for any other method.

The strongest parametric competitor is RaLM-CP, reported at 43.4 F1, which we strengthen with an experience-replay mix of unchanged facts. This gives a fair "continued pretraining done properly" comparison. It receives the same number of tokens and update steps per month as our sleep phase.

The key ablation is full regeneration. At each sleep, all compositions are discarded and regenerated from the current fact store, with no provenance. This is the accuracy upper bound for our approach. We also run a provenance-free reconsolidation ablation, which retrains a random set of compositions of the same size as the provenance-targeted set. It isolates whether targeting matters.

A third ablation removes the explicit disfavoring of old answers. A fourth removes the gate, to check whether the instability seen in When to Learn and When to Stop reappears.

All trained methods share the base model, the data snapshots, a fixed per-month GPU-hour budget (except full regeneration, whose cost we report), and an equal hyperparameter search budget. Results are averaged over at least three seeds, with variance reported.

What counts as success. The proposal succeeds if four things hold together:

  • flip-item accuracy is significantly above both replay-strengthened continued pretraining and RiLM;
  • flip-item accuracy is within a few points of full regeneration;
  • reconsolidation compute is a small fraction of full regeneration's;
  • performance on Unchanged knowledge does not fall below the frozen model.

Why not just…?

Why not just use retrieval and never touch the weights? RiLM already beats continued pretraining on GrowOVER, so this is a real threat. Our bet is that retrieval for a two-step question must find both hops, and the question text names only the first entity. Retrieval may therefore surface the changed fact without connecting it to the composition. The flip split of GrowOVER-Hop tests this directly. If RiLM matches us there, the parametric story loses, and we would instead study sleep as a complement that supplies retrieval with better intermediate entities.

Why not just keep training on the new text with a replay buffer? This teaches the changed single-hop fact. Nothing in it removes an already-composed answer stored elsewhere in the weights, which is exactly the rigidity seen in humans after a value flip (Memory reactivation during rest forms shortcuts in a cognitive map). The stale composition rate, conditioned on knowing the new hop, measures this failure. If it is near zero for replay-strengthened continued pretraining, our premise is wrong and we will report that.

Why not just regenerate every composition each cycle? This is the most likely way to match our accuracy, and we treat it as the oracle rather than a straw man. Its cost grows with the whole fact store, while reconsolidation grows with the number of changed facts times their dependents. We will plot accuracy against compute per month for both. If full regeneration is cheap at GrowOVER's scale, we will say so and rest the claim on scaling across snapshot sizes.

Proposed Experiments

The first experiment checks that the phenomenon exists in models. We compare models with and without self-generated dreaming on stable and flip items. The prediction, transferred from the human results, is that dreaming raises accuracy on stable two-step questions and raises the stale composition rate on flip items. If dreaming produces no rigidity, reconsolidation has nothing to fix, and the paper becomes a negative result about the human analogy.

The second experiment is the main comparison over the 2023 snapshots. We report:

  • GrowOVER-QA F1 by Unchanged, Changed, and New;
  • GrowOVER-Hop accuracy by stable, flip, and new;
  • the stale composition rate;
  • GPU-hours per sleep;
  • retention on Unchanged knowledge from earlier months.

A general-capability check on a fixed slice of LiveBench's programmatically graded tasks will detect whether sleep damages reasoning. We have found no reported LiveBench number on Lacuna for a 7B Llama2 model, so we will measure the frozen model ourselves as the reference.

The third experiment tests propagation depth. We extend GrowOVER-Hop to three-step chains and ask whether provenance-based invalidation, applied recursively through compositions built on other compositions, keeps the stale rate flat as chains lengthen. We expect full regeneration's cost and continued pretraining's stale rate both to grow with chain length.

The decisive pattern is three results together. Provenance-targeted reconsolidation must beat the random-target ablation on flip items at equal compute. It must stay close to full regeneration. And it must lower the stale composition rate below RiLM's. If all three hold, we will have shown that biological consolidation offers a usable mechanism for evolving benchmarks. Shortcuts are worth forming only if a system also remembers what they were built from.