Evidence Arrival Time: Why the Language of a Document Changes How Much a Model Believes It

When a language model reads a retrieved document that contradicts what it learned in pretraining, it has to decide whom to believe: the page in front of it, or its own memory. Studies of multilingual models show that this decision depends on the language of the document. In some languages models are stubborn, and in others they are easily persuaded, and the pattern does not simply follow how much training data each language has (When Abundance Conceals Weakness: Knowledge Conflict in Multilingual Models). This note proposes a single mechanism behind that pattern. Multilingual models translate their inputs internally into an English-like "concept space" in their middle layers. The circuitry that settles memory-versus-context conflicts also sits in those middle layers. A document in a language the model pivots slowly arrives in that shared space late, after the model has partly committed to an answer. English text, including irrelevant English text, arrives early and so receives more authority than its content deserves.

Put in two sentences: the authority a document gets is set partly by when its meaning becomes readable in the model's internal pivot space, relative to the layers where conflicts are resolved. If that is true, we can measure each language's arrival delay, predict stubbornness and distractor susceptibility from it, and correct it by placing a contextual-faithfulness intervention at a layer chosen per language, instead of translating every document first.

Core Claim

Two established findings line up in a way no one has tested. Probes of the residual stream (the running vector each layer adds to) in Llama3-8B detect memory-context conflict starting near layer 8, with about 90% probe accuracy at layer 14, and the decision becomes readable at layers 16–17 with AUROC above 0.85 (Analysing the Residual Stream of Language Models Under Knowledge Conflicts). Separately, the logit lens (decoding intermediate hidden states through the output vocabulary) shows that non-English inputs pass through an English-dominated middle phase before being converted to the target language at the end (Do Llamas Work in English? On the Latent Language of Multilingual Transformers; LinguaMap: Which Layers of LLMs Speak Your Language and How to Tune Them?). Both studies are monolingual or closed-book, and neither asks whether the pivot and the arbitration window overlap.

We claim they do, and that a document's language-specific arrival layer, defined below, predicts three behaviors. First, the Persuasion Rate under counterfactual evidence falls as arrival is delayed. Second, distractors in the model's dominant internal language pull answers more than equally irrelevant distractors in other languages. Third, in models whose internal pivot language is not English, the advantage moves to that language (What Language Do Non-English-Centric Large Language Models Think in?). The third prediction is the sharpest falsifier. A pure "resource abundance" account, which is one of the two paths CLEAR proposes, predicts no such shift.

Let $h_\ell^{(L)}$ be the residual state at the answer position when the evidence is written in language $L$. The arrival layer is

$$a(L) = \min \{\ell : \Pr_{\text{lens}}(y_{\text{ctx}} \mid h_\ell^{(L)}) > \tau \},$$

the first layer at which the logit lens assigns probability above a threshold $\tau$ to the answer the context asserts, $y_{\text{ctx}}$, in any surface language. The per-language delay is $\Delta(L) = a(L) - a(\text{en})$.

Research Program

The first component is a factorial test suite. We start from CLEAR's parallel ConflictQA-PopQA (898 factual items) and ConflictQA-StrategyQA (1,000 reasoning items) in 10 languages. We vary three things independently: the query language, the language of the counterfactual evidence, and the language of the added distractors. Distractors follow KnowQA's irrelevant and partly-irrelevant scenarios (Resisting Contextual Interference in RAG via Parametric-Knowledge Reinforcement) and are translated so that their content is held fixed. Every conflict item has a matched "correct evidence" item, so we can tell whether a method makes the model more faithful or just more gullible. This crossing of conflict and multilingual noise is exactly what CLEAR leaves open. CLEAR measured behavior under clean conflicting evidence and did no internal analysis. Linguistic Nepotism found an English preference in citation choices (Linguistic Nepotism: Trading-off Quality for Language Preference in Multilingual RAG), but it never set the document against the model's memory.

The second component is measurement on Llama3-8B, Qwen2.5-7B, and one Chinese-centric model. Llama3-8B has existing conflict-probe results to replicate. Qwen2.5-7B is the model Knowledgeable-R1 was tested on. For each model we compute $a(L)$, and we train conflict and decision probes on English and test them on every language, following the residual-stream study. If the probes transfer but trigger later layers in slower languages, that supports a shared arbitration circuit reached late. If they do not transfer at all, the claim of a shared circuit fails. We also apply the head-attribution method of Cutting Off the Head Ends the Conflict: A Mechanism for Interpreting and Mitigating Knowledge Conflicts in Language Models to ask whether the same "context heads" serve all evidence languages. This mirrors the shared-circuit finding for closed-book multi-hop reasoning in Bridging the Language Gap: Uncovering and Aligning Shared Circuits for Multi-Hop Reasoning in Multilingual LLMs, now carried into the setting with retrieved context.

The causal test is cross-language activation patching. We take a non-English evidence span and overwrite its residual states, at a chosen layer $\ell$, with the states from the parallel English evidence. If the arrival-time account is right, the Persuasion Rate should recover when the patch is applied before the arbitration window and should not recover when it is applied after. A comprehension account predicts something different: recovery should depend only on whether the patch happens at all.

The third component is the intervention, language-calibrated steering. We build one contextual-faithfulness steering vector from English data with the ContextFocus recipe (ContextFocus: Activation Steering for Contextual Faithfulness in Large Language Models). ContextFocus adds this vector at a fixed layer. We instead inject it at layer $\ell^* + \Delta(L)$, where $\ell^*$ is the best English layer. We also add a weak opposing component at the positions of distractors written in the dominant language. The intervention uses one forward pass, no fine-tuning, and one calibration constant per language.

Baselines

The naive reference is plain retrieval-augmented prompting with no intervention. It gives the raw per-language Stubborn Rate and Persuasion Rate under each combination of evidence and distractor language. For context, CLEAR reports averages of 13.4% Stubborn Rate and 81.0% Persuasion Rate on ConflictQA-PopQA, and 28.4% Stubborn Rate on ConflictQA-StrategyQA (When Abundance Conceals Weakness: Knowledge Conflict in Multilingual Models). We will re-measure these on our models, not assume them.

The strongest accuracy competitor is self-translate (Do Multilingual Language Models Think Better in English?). The model first translates the query and documents into English, using FLORES-200 few-shot prompts, and then answers. It is the strongest competitor because it removes the arrival delay by brute force. Its reported effect is a gain on LLaMA-30B from 43.1% to 46.6% average accuracy across five tasks under direct inference versus self-translate, but on reasoning tasks, not under conflict. We run it with the same models and items and report it at matched accuracy against wall-clock time and generated tokens. Beating it means matching its language-parity gains within noise at roughly half its inference cost, or exceeding it on items with non-English distractors, where translating everything into English could make nepotism worse.

The strongest trained competitor is Knowledgeable-R1. Its reported result is up to +30.47% over standard prompting on counterfactual KnowQA with Qwen2.5-7B, with about 80% accuracy kept on correct context. We retrain it on our multilingual mixture with the same item budget. If it closes the language gap as well as our method, the timing story still holds as diagnosis, but the intervention adds no practical value, and we would say so.

The key ablation is fixed-layer ContextFocus with the English vector at $\ell^*$ for every language. No reported ContextFocus number was found on Lacuna; the paper gives only qualitative gains on ConFiQA. The only difference between this ablation and our method is the per-language layer shift. If fixed-layer steering does as well, the arrival-time ingredient is unnecessary. We also include Soft Constrained Decoding (Language Drift in Multilingual Retrieval-Augmented Generation: Characterization and Decoding-Time Mitigation), which raised HotpotQA ZH-EN ROUGE from 0.182 to 0.306. It controls the output language and should be orthogonal to arbitration, which makes it a check that our gains are not just better output-language consistency.

Why not just…?

Why not just translate every document into English before answering? This is the real threat, because it attacks the delay directly. It has three costs. It doubles inference, it introduces translation errors that are hardest to catch in low-resource languages, and by moving all evidence into English it may hand English distractors even more pull. The matched-cost comparison and the non-English-distractor condition settle whether it dominates. If it wins on both, our contribution narrows to explaining why it works.

Why not just apply the English ContextFocus vector unchanged? If arbitration happens in a shared space, the English vector should already transfer, and part of our hypothesis predicts exactly that. The open question is placement. The fixed-layer ablation tests this directly, and a flat result across $\Delta(L)$ would falsify our design.

Why not just train Knowledgeable-R1 on multilingual data? Reinforcement learning does transfer across languages for closed-book reasoning (Beyond English-Centric Training: How Reinforcement Learning Improves Cross-Lingual Reasoning in LLMs), so this may work. It needs training per model, though, and its fallback assumes the memorized fact is correct. We run it as a baseline and report the result honestly.

Isn't "late arrival" just poor comprehension of the language? We control for this with no-conflict extraction: can the model answer from the same evidence when memory does not object? If extraction accuracy is high but the Persuasion Rate is low, the cause is timing, not comprehension. We restrict the main analysis to languages that pass this check.

Decisive Checks

The account survives only if four things hold. $\Delta(L)$ must correlate with the per-language Persuasion Rate across the 10 CLEAR languages, after controlling for extraction accuracy. Cross-language patching must recover persuasion only when applied before the probe-identified decision layers. The advantage of dominant-language distractors must move to Chinese in the Chinese-centric model. Calibrated steering must beat fixed-layer steering on the persuasion gap across languages without lowering correct-evidence accuracy. If the third check fails, the resource-abundance account wins, and we report that outcome.