The Switch Matrix Updated Aug 2026
Handoff drift in multi-turn LLM systems

What happens when one model finishes another's conversation

On 7 August 2025, OpenAI launched GPT‑5 and removed GPT‑4o from the model picker. Existing conversations did not reset. A thread with hundreds of 4o‑authored turns simply received its next turn from a different model — across a product with roughly 700 million weekly users.

Nobody measured what that does. Seven months later, a team at NatWest AI Research and UCL built a benchmark for exactly this: a switch matrix, where one model writes the early turns and a different model writes the last one.

Analysis  On the GPT‑5 framing above

The rollout is our framing, not the paper's — the paper never mentions it. It is a useful illustration and not evidence for these results. The rollout changed the model, its personality, its rate limits and its routing all at once, with no counterfactual; the study below changes exactly one thing and holds the rest fixed. Treat the rollout as the reason to care, and the matrices as the actual finding.

The setup

A handoff is a structured distribution shift: the model writing the final turn is conditioning on a dialogue it did not write.

NO-SWITCH BASELINE (B → B) turns 1 … L−1 Model B final turn Model B score sB→B SWITCH CELL (A → B) turns 1 … L−1 Model A · the prefix final turn Model B · the suffix score sA→B Δ A→B = s A→B − s B→B measured per episode, on the same task negative = the foreign prefix hurt B
Method
Models
9
each used as both prefix and suffix — a 9×9 grid
Episodes
200
randomly sampled, per benchmark
Switch policy
Final turn
the suffix writes exactly one turn in every cell
Uncertainty
Paired BCa
1,000 bootstrap resamples, paired within episode
Settings
temp 0
2,048 max output tokens; reasoning effort and verbosity low where supported

The switch matrices

Rows are the prefix model that wrote the early turns. Columns are the suffix model that wrote the final one. Each cell is ΔA→B against that suffix model's own no-switch baseline — blue means the foreign prefix helped, red means it hurt. Hover any cell for the exact value, its significance and the underlying cell mean. The hatched diagonal is the baseline itself.

Suffix model — writes the final turn →

← scroll the grid sideways · row labels stay fixed

· stars = bootstrapped CI excludes 0 at 90% (*), 95% (**), 99% (***)

What the grid says

Three findings, in the paper's own terms.

One handoff is enough. Of the 72 off-diagonal switches, 16 on CoQA (22%) and 18 on Multi-IF (25%) are significant at the 95% level — from a single final turn written by a different model.

It is directional, not symmetric. A handoff that harms B after A may be neutral or beneficial in reverse. Handoff robustness is a property of the ordered pair and the dialogue regime the prefix establishes, not of model quality. Within-family switches among the GPT‑5 variants sit near zero on CoQA, which the authors read as shared conventions reducing mismatch.

It cuts both ways. The largest Multi-IF effects are positive: a stronger prefix can stabilise a compliant output regime and lift a weaker suffix. The failures look like behavioural anchoring — the suffix continuing, or failing to override, a protocol the prefix set — rather than lost capability.

The extremes

The ten largest single-handoff effects in each benchmark. Positive means the foreign prefix improved the suffix model's final answer. These are the tails, not the typical case — see the note below the chart.

Two numbers per model Analysis

The paper shows drift is largely additive: ΔA→B ≈ μ + αA + βB, where α is prefix influence (how much a model's dialogue style helps or hurts whoever continues it) and β is suffix susceptibility (how much a model gains or loses when handed a foreign history). That compresses a 9×9 grid into 18 numbers you can monitor.

α · prefix influence — higher means this model's history helps whoever writes next

β · suffix susceptibility — positive means this model tends to gain from a foreign prefix; negative means it degrades

How these values were produced

The paper publishes α and β as bar charts and states six β values in its text, but does not print the full set. We refit the paper's own model — equation 1, ordinary least squares on the 72 off-diagonal cells with sum-to-zero constraints — to the published Δ tables. The fit reproduces the paper's stated R² (we get ) and all six stated β values to within 0.0007. Rank order matches the paper's figures in three of four charts; the exception is CoQA β, where GPT‑5.2 and Gemini‑2.5‑Flash differ by less than 0.0001 and swap places — a rounding artefact of fitting on three-decimal published values rather than raw episode scores. Treat the ordering as sound and the last decimal as ours, not theirs.

What to do about it

The paper's operational recommendations.

Log the authoring model per turn, and monitor the first turn after a switch specifically — a handoff is a discrete change point, and averaging across a session hides it.

Run a handoff regression before upgrades or fallbacks. Replay historical prefixes through the candidate suffix model to estimate ΔA→B, and flag risky pairs or spot uplifts before they reach users.

Track α and β as a compressed view, then prioritise deeper evaluation for the pairs with large residuals. Gate routing when deviation from the no-switch baseline is large; the authors suggest injecting a short handoff instruction as a simple mitigation.

The paper measures rather than fixes. It names explicit handoff summaries, lightweight adapters and continuity-aware routing policies as future work, and does not test them.

Where it landed Analysis

This timeline is our assembly of public, dated artefacts. The connections drawn between them are ours.

3 Mar 2026
The paper is posted
arXiv:2603.03111, later accepted at the ICLR 2026 CAO Workshop.
5 Mar 2026
vLLM Semantic Router files the bug
Issue #1439: routing is evaluated per request with no session affinity, so a short follow-up like “commit it” can land on a far smaller model than the one that wrote the code being discussed.
6 Mar 2026
… and proposes the fix
Issue #1458 adds Conversational Routing Momentum — an asymmetric low-pass filter borrowed from audio compressor attack/release dynamics — shipped in the v0.3 “Themis” milestone.
7 Aug 2026
Lemonade SDK proposes the same thing, and cites the paper
Issue #2956 proposes stateless hysteresis for its router, quoting this paper's −8 to +13 pp range as evidence that switching itself degrades outcomes — not merely that reloading a model is slow.
Read this carefully

Two independent routing projects converged on the same mitigation shape within months. But the vLLM Semantic Router issues do not cite this paper, and they were filed two and three days after it was posted — suggestive timing, no demonstrated causal link, and we are not claiming one. Formal academic uptake is meanwhile close to zero: OpenAlex records no citations and Google Scholar one, as of August 2026. The paper's visible afterlife so far is as engineering justification, not as a benchmark anyone has rerun.

v1.0