A Second Pass, a Different Query

TL;DR for operators

An enterprise AI assistant may retrieve two related facts correctly when asked separately, yet fail to combine them into one answer. Research on looped language models shows that later processing passes can repair this specific failure. Transplanting later-loop attention queries into an earlier pass rescued 89.8% and 96.9% of selected synthetic errors in two Ouro models, substantially outperforming control interventions.

These figures are conditional rescue rates, not overall accuracy improvements. The experiments suggest that updated attention queries help direct intermediate retrieval, but natural-language results are weaker and do not establish the same mechanism.

For enterprise question-answering teams, the immediate value is a better diagnostic approach: test whether failures arise from individual fact retrieval, combining facts, or reliance on document position. The study does not establish a faster inference method or a production-ready reliability improvement.

Two correct lookups do not guarantee a correct combined answer

Consider an assistant answering a customer-service question: Which account manager is responsible for this customer’s account?

The answer requires two lookups. First, identify the account belonging to the customer. Then, retrieve the manager assigned to that account. An assistant may answer either question correctly in isolation but fail when both relationships must be combined.

Additional model computation sometimes resolves this failure. The difficult question is what changes during that computation.

Ziyu Mu and Zehua Wang investigate this problem in Cross-Loop Query Handoff: How Later Loops Redirect Two-Hop Retrieval in the Ouro Looped Language Models.¹

The researchers examine two pretrained Ouro models, which repeatedly process information through the same transformer layers. Each recurrent pass, or loop, can update the model’s internal representations.

On synthetic questions requiring two successive lookups, accuracy changes substantially between loops 1 and 3:

Source: Section 3 and Appendix C, Table 7; 400 synthetic scenes per model.

The corresponding one-hop questions already achieve between 95.50% and 99.50% accuracy at loop 1. After subtracting the larger one-hop improvement within each scene, the researchers report composition-specific gains of 28.7 and 33.0 percentage points.

The models are therefore improving at combining facts, beyond their already strong ability to retrieve individual facts.

That establishes a behavior worth investigating. It does not explain how the later loops produce the improvement.

Later attention queries rescue selected earlier failures

An attention query helps determine which information a model retrieves from its available representations. In a two-hop question, successful computation must use the first retrieved fact to identify the appropriate target for the second lookup.

The researchers test whether attention queries from a later loop can supply this missing retrieval guidance.

They first run the model normally and cache its loop-3 queries. They then replace the loop-1 queries at the final prompt position with those cached queries, using corresponding layers and attention heads throughout the model. Earlier prompt positions retain their native loop-1 states.

This is an artificial intervention. It transfers a collection of queries across all layers and heads, rather than inserting one intermediate vector.

The main experiment selects questions that the model answers incorrectly at loop 1 but correctly at loop 3.

Source: Section 4, Table 2; 600 synthetic two-hop scenes per model. Rescue is conditional on failure at loop 1 and success at loop 3.

The same-example advantage is 66.1 percentage points for Ouro-1.4B, with a 95% confidence interval of [60.0, 72.2], and 76.2 points for Ouro-2.6B, with an interval of [70.4, 81.6].

The shuffled-donor control matters. Replacing an early query with an arbitrary later query does not reproduce the effect. The improvement depends substantially on information specific to the original question.

The synthetic effect also replicates on newly sampled scenes. For Ouro-2.6B, a greedy-generation test rescues 89.6% of selected failures, compared with 20.3% under shuffled queries. For Ouro-1.4B, however, the intervention changes candidate rankings without producing a color answer in greedy generation.

The central finding is consequently precise: later-loop attention queries contain transferable information that can repair many selected earlier-loop retrieval errors under the tested intervention.

It is not a claim that nearly every two-hop question becomes correct, or that the model has learned an inexpensive shortcut.

Queries help direct retrieval, but the mechanism is not exclusively semantic

The researchers use additional interventions to determine what the transferred queries actually do.

In a crossed-swap experiment, they transplant queries between different synthetic scenes. The candidate outcomes distinguish retrieving through the donor’s intermediate bridge entity, copying the donor’s original answer, and following a fact-line position.

At loop 1, the recipient selects the answer associated with the donor’s bridge entity in 45.0% of Ouro-1.4B cases and 48.0% of Ouro-2.6B cases. Direct donor-answer copying accounts for only 13.3% and 8.2%, respectively.

This favors bridge-conditioned retrieval over straightforward answer copying.

However, the bridge-consistent outcome remains below half of predictions. In Ouro-2.6B, another 34.8% follow the donor-answer line position. Because entity identity and line position are not independently randomized, the experiment cannot isolate purely entity-based routing (Section 5, Table 3).

Two further tests sharpen the interpretation.

First, a separate converse transplantation experiment replaces late keys and values associated with the relevant facts instead of replacing queries. Key-value transplantation rescues only 26.2% and 33.2% of selected errors, compared with 86.7% and 98.2% for query transplantation (Section 5.1, Table 4).

Second, the researchers freeze attention components at their loop-1 values during later passes. Freezing final-position queries eliminates the native two-hop loop gain in both models. Matched questions that explicitly provide the bridge entity suffer 17.0 and 30.3 percentage points less damage, respectively, suggesting a composition-specific role for query updates (Section 6, Table 5 and matched bridge-given controls).

These tests are complementary rather than interchangeable. Transplantation tests whether a later representation can help an earlier computation; freezing tests dependence on native updates; crossed swaps investigate the information guiding retrieval.

Freezing used-fact values also severely disrupts performance. In Ouro-2.6B, jointly freezing keys and values is less damaging than freezing values alone, suggesting that compatibility between internal representations may matter.

The resulting mechanism is partial: evolving queries participate in directing retrieval, while changing fact representations also contribute. The study does not locate where the bridge entity is computed or demonstrate that one compact representation carries the complete intermediate answer.

Natural-language questions expose a narrower effect

The strongest findings come from synthetic scenes with four people, four boxes, and four colors. These environments make competing retrieval explanations measurable, but their structure is considerably simpler than ordinary documents.

On natural multi-hop questions from 2WikiMultiHopQA and MuSiQue, query rescue is substantially weaker.

For Ouro-2.6B, the original all-layer intervention rescues 33.6% of 134 selected failures, compared with 19.4% for shuffled queries. An independent evaluation on unused 2WikiMultiHopQA items finds a smaller, replicated advantage of 8.2 percentage points.

Ouro-1.4B shows no statistically supported natural-question rescue: 7.4% versus 6.3%, a difference of 1.1 points with a confidence interval spanning zero (Section 7, Table 6).

A follow-up intervention restricted to the final quarter of Ouro-2.6B’s layers produces larger advantages of 29.9 points initially and 25.5 points in replication. This suggests that intervention placement matters in that model and setting, but it does not establish a general layer-selection rule.

More decisively, the natural-question crossed-swap mechanism test is inconclusive. Across 711 eligible pairs, donor queries produce the bridge-associated answer in 27.7% of cases, versus 26.2% with unrelated queries. The 1.5-point difference has a 95% interval of [-0.1, 3.2].

Thus, a smaller query-rescue effect transfers to natural questions in the larger model, but the synthetic explanation involving bridge-conditioned redirection is not independently confirmed.

Alternative prompt wording, distractors, and larger synthetic scenes also weaken some findings. Ouro-1.4B even loses the required composition-specific improvement under certain configurations (Section 8).

These are robustness and generalization tests, not additional demonstrations of the original mechanism.

What enterprise question-answering teams can test

Cognaptus inference: The study offers a way to diagnose multi-step retrieval failures more precisely, rather than an immediately deployable improvement to enterprise question answering.

For a team building an assistant over customer records, financial documents, or internal knowledge bases, the evaluation can distinguish several failure modes:

1. Individual retrieval: Can the assistant independently identify the customer account and its manager?
2. Composition: Can it connect the two facts without receiving the account identifier explicitly?
3. Bridge-given control: Does providing the intermediate account identifier remove the failure?
4. Routing sensitivity: Does the answer change when relevant records are reordered, distractors are introduced, or equivalent questions are reworded?
5. Full-workflow accuracy: Do improvements survive unrestricted answer generation and realistic document retrieval?

These tests inform different decisions. If individual lookups fail, retrieval quality or source coverage may need attention. If individual lookups succeed but their combination fails, the team has stronger grounds to investigate intermediate retrieval behavior, explicit multi-stage workflows, or additional model computation.

For developers with access to model internals, query transplantation and matched freezing offer more specialized diagnostics. Enterprise teams using closed models can still implement the behavioral controls without claiming to observe the same internal mechanism.

The boundary is significant: these are proposed applications of the paper’s methodology, not experimentally validated enterprise interventions.

The missing step between mechanistic rescue and cheaper inference

A successful transplant might suggest that loop 1 could achieve loop-3 performance without completing the intervening computation.

The experiment does not establish this.

The late queries must first be generated through the donor’s later loops. Transplanting them back into loop 1 therefore does not remove their computation cost. A practical early-exit method would require an independently demonstrated way to produce sufficiently informative queries at lower cost.

There are additional architectural limits. The study examines two related Ouro checkpoints; its limited dense-transformer controls do not match effective depth closely enough to attribute the effect specifically to recurrence. Patched internal states are also artificial combinations that may differ from states encountered during normal inference.

For inference engineers deciding whether to invest in adaptive looping, the paper identifies a possible diagnostic target, not a verified latency optimization.

For enterprise reliability teams, its contribution is more immediate: final-answer improvement alone cannot identify whether a model has learned to combine the intended facts, exploit their positions, or rely on another representation-dependent behavior.

The study makes that distinction experimentally tractable in a controlled setting. Extending it to realistic documents remains the next necessary validation before changing operational safeguards or inference architecture.

Cognaptus: Automate the Present, Incubate the Future.