Consider a model reading a situation described in Amharic or Hindi that never names a feeling, such as a son skipping a family gathering to finish work. It has to infer what someone from that culture would feel, and the same event can mean guilt in one country and nothing at all in another. Two benchmarks on this kind of implicit emotion found something odd. Large language models do better when the scenario is posed in English with the country named than when it is posed in the native language. CULEMO reports that GPT-4's accuracy on Indian scenarios rises by 21% under English-with-country prompting relative to Hindi. Tears or Cheers? (CEDAR) reports similar native-language drops. Separately, interpretability work shows that multilingual models internally map non-English inputs onto English-like representations in their middle layers: Can you map it to English? patched English activations into layers 14–25 of failing non-English passes and flipped the answers to correct.
We propose to connect these observations with a single testable account. The culture-specific way of appraising an event survives in the model's early layers of a native-language input, but it is weakened or overwritten when the input is mapped onto the internal English representation. If that is right, a small "culture direction" can put it back after the pivot layers. The direction would be computed from a handful of native-annotated examples, which is the cold-start regime for a new culture, and added to the model's hidden state at inference time. This requires no translation and no fine-tuning. The core idea fits in two sentences: models lose cultural emotion cues when they internally route foreign-language text through English, and we measure where the loss happens and add the missing cue back as a vector estimated from a few labeled examples.
Core Claim
The claim has two parts, and each can be falsified on its own. The first is diagnostic. On scenarios whose correct emotion label differs by culture, a linear probe can decode the culture-appropriate label from native-language hidden states in early layers. That decodability should drop across the middle-layer band where cross-lingual alignment happens, while the English-with-country-tag input keeps it. The second part is interventional. A culture direction, estimated from $k$ labeled scenarios and added to native-language query passes just after that band, should recover most of the English-with-tag advantage. It should also beat $k$-shot in-context prompting that uses the same $k$ labels.
The mechanism combines two existing results. In-Context Learning Creates Task Vectors showed that a set of demonstrations is compressed into one hidden-state vector at the last demonstration token. Patching that vector into a query-only pass recovers 80–90% of in-context accuracy, compared with 10–20% without it. Steering Llama 2 via Contrastive Activation Addition showed that differences of activations over contrastive prompt pairs steer behavior additively, with negligible effect on MMLU. Neither paper touched subjective, culture-dependent, or cross-lingual tasks: the task-vector paper used single-token lookups like capitals and antonyms, and the steering paper used behaviors like sycophancy. Our contribution is to build the contrast from identical inputs that carry different cultural labels, so that the difference isolates cultural appraisal rather than content. We then place the vector relative to the English pivot identified by the cross-lingual patching work.
Construction of the Culture Direction
CuLEmo makes the contrast clean because each of its 400 situational questions is annotated by at least five native speakers in each of six countries (USA, Mexico, Germany, UAE, Ethiopia, India). For a demonstration set $D$ of $k$ scenarios, let $\theta_L(D, c)$ be the hidden state at layer $L$ at the final demonstration token, where each scenario in $D$ is paired with culture $c$'s majority label. The culture direction for culture $c$ is
$$\delta_{c,L} = \mathbb{E}_{D}\left[\theta_L(D, c) - \theta_L(D, \bar{c})\right],$$
where $\bar{c}$ is a reference labeling (the pooled majority across the other five cultures) and the expectation averages over random demonstration subsets and orderings. The scenarios on both sides of the difference are the same, so anything shared, such as topic, language, and the generic "label an emotion" task, cancels out. What remains is how culture $c$ maps events to emotions.
At test time, a native-language scenario with no demonstrations is run through the model. We add $\alpha\,\delta_{c,L'}$ to the residual stream at the last token, at layer $L'$, with $\alpha$ a scalar strength. The injection layer $L'$ and the extraction layer $L$ are swept, with particular attention to positions before, inside, and after the pivot band. The prediction is specific: injection before the band should be partly washed out by the English mapping, and injection after it should work best.
We build a second variant from native-language demonstrations and a third from English-translated demonstrations. The third variant tests whether a direction learned in English space carries over to native passes. Such a direction would be the cheapest to obtain for a new culture, because English translations of scenarios are easier to annotate at scale.
Research Program
The work has three stages.
The first stage is the diagnostic. For open-weight models that allow activation access, we train linear probes at every layer. The models are the Mistral-7B and Qwen1.5 families studied in Getting More from Less, a Llama model, Aya (which CEDAR found most stable across languages), and Gemma3-27B-It. Each probe predicts culture $c$'s label from the hidden state of a scenario presented in three conditions: in language $l$, in English with a country tag, and in English without one. We restrict the probe analysis to "divergent" scenarios, where at least two cultures' majority labels differ, because only there does the culture signal carry information. We also compute the instance-level cross-lingual alignment scores used in Can you map it to English?. This tells us whether the scenarios where the native input aligns most tightly to English are the ones where cultural decodability collapses most, which would directly link the two literatures' findings.
The second stage is the intervention and its cold-start curve. We sweep $k \in \{4, 8, 16, 32, 64\}$ labeled scenarios per culture and split by CuLEmo's ten societal categories: directions are estimated on five categories and tested on the other five, so no scenario content leaks between estimation and test. This is the realistic cold-start setting, in which a deployer entering a new culture can afford a few dozen native annotations but not a dataset.
The third stage is transfer. We test whether a direction estimated on CuLEmo transfers to a different benchmark in the same language and culture: the India direction to CEDAR Hindi, the UAE direction to CEDAR Arabic, and the Mexico direction to CEDAR Spanish. CEDAR's text-only split has 1,166 samples per language over 14 fine-grained emotions, and its emotion words were scrubbed out. Its label space differs from CuLEmo's, so for this stage we reuse the direction only and read out CEDAR labels through the model's normal zero-shot prompt. This asks whether the vector encodes an appraisal tendency rather than a label-set shortcut.
We deliberately do not pursue synthetic data generation in the style of LUAS, the dialogue-state-tracking work that bootstraps new domains with GPT-4 simulated conversations. For cultural appraisal, a simulator would generate labels from the same English-centric priors we are trying to correct, so the problem would be circular. LUAS also reports that its synthetic dialogues lose human irregularity even in an explicit-slot setting. Implicit emotion detection in dialogue (ProEmoTrans) is the natural extension once the single-scenario case works, but it is not part of this proposal.
Baselines
The do-nothing reference is zero-shot prompting in the native language with no country information. This is the condition CuLEmo and CEDAR show to be weak. The only absolute figure in the file is Gemma3-27B-It at 30.46% on CEDAR's Japanese multimodal split. For open 7B models on CuLEmo's text scenarios, no reported result was found on Lacuna, so we will measure this baseline ourselves.
The strongest existing reference is English prompting with an explicit country tag, applied to a machine translation of the native scenario. This is exactly the condition behind CuLEmo's reported 21% gain for GPT-4 on Indian culture over Hindi-only prompting. The number is relative, and no absolute CuLEmo accuracy was found in the file, so we re-run this condition on every model with the same translation system for all languages. Our method beats it only if it matches or exceeds its accuracy without translation. If the direction closes most of the native-to-English gap but not all of it, we report that as partial recovery rather than a win.
The main in-context baseline is $k$-shot prompting with the same $k$ labeled scenarios, in the native language and with the country tag in the native language. It is run with every ordering and subset that the direction averages over. We report mean and variance, since the instability that A Systematic Survey of Automatic Prompt Optimization Techniques documents is part of what a vector average should remove. As a supervised reference we also fine-tune a LoRA adapter on the same $k$ examples with a matched hyperparameter search budget, and include mT0 from Crosslingual Generalization through Multitask Finetuning as an instruction-tuned multilingual model.
Three ablations isolate the key ingredient. The first uses a shuffled-culture direction, built from labels drawn from the wrong culture. If it helps as much as the true direction, we have only found a generic "attend to emotion" direction. The second uses a plain task vector without the contrast, which tests whether subtracting the reference labeling matters. The third injects before the pivot band instead of after it, which tests the mechanistic story itself.
Why not just…?
Why not just translate into English and add the country tag? This is the real threat, since it is the very condition that produced the paradox. Zero-shot Reading Comprehension by Cross-lingual Transfer Learning found that translation pipelines underperformed direct multilingual transfer, and CEDAR reports culturally loaded details being lost in the English pivot. However, both results come from different tasks or from diagnostics rather than a controlled comparison. The baseline above settles the question per language. If translate-plus-tag wins everywhere, the intervention's claim narrows to a translation-free alternative, and the probe study stands as the main result.
Why not just put the $k$ examples in the prompt? Few-shot prompting needs no model access and may be enough. The case for the vector is twofold. First, averaging over subsets removes ordering sensitivity. Second, the task-vector paper's conflict experiments show that patched vectors override competing demonstration content, which matters when culture labels contradict the model's English-centric prior. The matched $k$-shot comparison, reported with its variance, is the direct test. If in-context prompting ties the vector at every $k$, we would conclude the vector adds only the lower inference cost.
Why not just name the country in the native-language prompt? This isolates whether the model simply lacks the culture cue. We predict that a native-language tag helps less than an English tag, because the tag itself passes through the same pivot. If a native tag closes the gap alone, the "lost at the pivot" story is wrong, and the probe results would show culture information intact in the middle layers.
Why not just fine-tune on the few examples? Fine-tuning a separate adapter per culture is heavier than storing one vector per culture, and with 4–64 examples it risks overfitting to CuLEmo's categories. The held-out-category split and the CEDAR transfer test measure exactly that overfitting.
Decisive Checks
We report, per model, per culture, and per value of $k$: accuracy against the native majority label; Jensen–Shannon divergence between the model's label distribution and the annotator distribution; and a culture-sensitivity score, defined as the fraction of divergent scenarios where the model's answer changes in the annotators' direction when the culture changes.
The proposal succeeds if three things hold together. The culture-label probe accuracy on native inputs must drop measurably across layers 14–25 relative to English-with-tag inputs. The post-pivot direction at $k \le 32$ must recover at least half of the gap between the native zero-shot and translate-plus-tag baselines, while beating matched $k$-shot prompting in mean and variance. And the shuffled-culture control must give no comparable gain. The account fails if the probe shows no drop in decodability, or if pre-pivot and post-pivot injections perform the same. Either outcome would mean the English Prompt Paradox arises somewhere other than the internal English mapping, which is itself worth knowing. The CEDAR transfer test decides whether the vectors encode cultural appraisal or only CuLEmo's label conventions.
References
- CULEMO: Cultural Lenses on Emotion - Benchmarking LLMs for Cross-Cultural Emotion Understanding
- Tears or Cheers? Benchmarking LLMs via Culturally Elicited Distinct Affective Responses
- Can you map it to {E}nglish? The Role of Cross-Lingual Alignment in the Multilingual Performance of {LLM}s
- In-Context Learning Creates Task Vectors
- Steering Llama 2 via Contrastive Activation Addition
- Getting More from Less: Large Language Models are Good Spontaneous Multilingual Learners
- Enhancing Dialogue State Tracking Models through LLM-backed User-Agents Simulation
- Emotion Transfer with Enhanced Prototype for Unseen Emotion Recognition in Conversation
- A Systematic Survey of Automatic Prompt Optimization Techniques
- Crosslingual Generalization through Multitask Finetuning
- Zero-shot Reading Comprehension by Cross-lingual Transfer Learning with Multi-lingual Language Representation Model