acceptodds
Under review as a conference paper at ICLR 2027

From State Behavior to Steerable Expression: Geopolitical Coordinates in Large Language Models

Abstract

Large language models can express diverse political perspectives. We investigate whether geopolitical information in their country representations extends to authentic diplomatic language and can guide changes in generated expression. We show that internal country knowledge can support both reading behavior-linked geopolitical information from authentic diplomatic speech and constructing interventions that guide generated expression. Across three instruction-tuned models, we extract attention-head activations from prompts placing named countries in diplomatic contexts. Linear probes map these activations to country positions estimated from United Nations voting records. Strong country-ordering signals are linearly readable from multiple attention heads spanning different layers in all three models. Readouts fitted to country-name features and labels matched to the speech period also predict country ordering from authentic speeches with speaker-country names masked, achieving Spearman correlations of 0.311–0.503 without fitting on speech features. Country-derived activation interventions shift the externally scored perspective of short diplomatic generations under fixed prompts and model weights, with matched random controls and an aggregate human audit supporting model-dependent effects. As a cross-model extension, source-model scores supervise separately fitted country readouts in another model, without true political labels in target fitting. Together, these findings connect observed state behavior, internal country knowledge, and diplomatic language, grounding the measurement and modulation of geopolitical expression in language models.

Primary area:
interpretability and explainable AI
Keywords:
large language models, representation interpretability, geopolitical perspectives

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_c5PFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper examines how the geopolitical information captured in country representations and authentic language speech representations -- extracted from attention-heads, using linear probes. The paper also demonstrates whether that can be transferred cross models, and whether that can be intervened.

The paper, using data from UN voting records, shows high correlation between country representations and their geopolitical positions, but less correlations when using speech representations. The paper also shows that it's possible to intervened the model generation output. Also, the linear probes can be transferred cross models.

Critical strengths

This is an interesting paper showing how LLMs capture geopolitical informations in their attention-heads. The paper wisely makes use of existing works for its purposes, showing evidence for the claims.

Critical weaknesses

The paper uses correlations to show whether the information of interested is captured, based on linear probes. Although this is quite standard in this line of research, it can be misleading.

  • First, linear probes can be very expressive (or not) depending a lot on the nature of the data. Here, the authentic speech has quite much lower correlation coefficient might be because its corresponding attention-head representations are less easy to exploited by the linear probes, compared to country name.
  • Second, using correlation coefficient \rho alone is pretty less informative. For instance, how could we say 0.8 is good (or bad). Therefore, a baseline could be helpful (e.g. using some vague information, such as location of the country of interest -- e.g. replace "China" with "Asia country").

It's unclear about the significance, importance of the contributions of the papers. It seems that the paper's technicality builds upon existing work in the literature. In terms of findings, how exactly the community can benefit from the findings?

Questions for authors

Are lambdas in equation 3 and 5 the same?

2 comments
AuthorsOpenReview ↗

Part 1/2: What Geography Explains and What the Speech Readout Measures

Thank you for the care you put into reviewing our paper and for suggesting a geographic baseline. In response, we added a geography-only predictor and checked whether the readouts distinguish countries within the same region. Your suggested baseline is informative: fine geographic groups alone yield ρ = 0.666. Even within Asia, however, the three models track differences in voting alignment among the 47 countries (ρ = 0.652–0.679).

1. New rebuttal analysis — a geographic baseline for interpreting ρ

We gave the baseline each country's actual UN M49 geographic group. For each held-out country, it predicted the mean voting ideal point of training countries in that group. We compared it with the original 64-head country-name readouts on the same 190 geographically matched countries, using the original five folds.

PredictorHeld-out Spearman ρ
Broad geographic region alone0.588
Finer geographic group alone0.666
Country-name readout, Qwen / Llama / Mistral0.857 / 0.839 / 0.838

The within-Asia comparison follows directly from your example of “an Asian country.” We also checked the result across the full cohort by subtracting each finer group's mean rank from both the predicted and observed ranks. The resulting correlations are 0.637 for Qwen, 0.599 for Llama, and 0.597 for Mistral. Thus geography explains some of the ordering, and the readouts recover additional differences among countries that share a geographic label.

2. Evidence in the submitted paper — the role of the probe on authentic speech

We agree that relevant information in speech may be harder to extract linearly. Section 4 fits and tunes the probe on country-name activations, then evaluates country rankings obtained from masked speeches. The lower speech correlation tells us how well this readout transfers. The size of the gap alone cannot tell us how much political information each representation contains.

Submitted Table 9 directly compares two linear readouts on the same speeches from the same 186 countries. Both exclude held-out countries from head selection and fitting. The only change in this comparison is whether we use the top head or average predictions from the top 64.

ModelSingle head, ρ64 heads, ρ95% CI for the increase
Qwen0.3290.535[0.073, 0.337]
Llama0.1650.501[0.139, 0.525]
Mistral0.1960.353[−0.012, 0.324]

This supports your point about accessibility. On identical speech inputs, the choice of readout makes a substantial difference for Qwen and Llama. The increase is less certain for Mistral. Submitted Table 7 tests a different question on country names: replacing ridge with a shallow MLP at the same linearly selected heads changes ρ by +0.005, −0.005, and −0.023, respectively. Whether a nonlinear decoder could recover more from authentic speech remains open.

3. Clarification of Equations (3) and (5)

The two λ values are set separately. Equation (3) fits a predictive readout on standardized features with an unpenalized intercept. The main country-name ensemble uses λ = 100, and the single-head speech analysis selects λ by inner cross-validation on training countries. Equation (5) constructs the steering direction from raw country-name activations with λ = 1 and no intercept. The intervention strength is α. We will distinguish the penalties in the notation and add the geographic baseline alongside the country-name results.

AuthorsOpenReview ↗

Part 2/2: Using Voting Records to Study Diplomatic Language and Compare Models

You asked what other researchers can gain from these findings. One practical use is to test whether a political coordinate learned from country identities responds to what a speaker actually advocates. A second is to use scores from one model to establish and independently evaluate a country readout in another.

1. Evidence in the submitted paper — from country names to diplomatic language

In Section 4, we fitted readouts on country-name activations using voting-derived targets from the speeches' reference period. We applied them to 1,710 authentic UN speech passages from 186 countries after masking specified speaker-country names. Country scores obtained by averaging passage predictions tracked the voting-based ordering (ρ = 0.503 for Qwen, 0.449 for Llama, and 0.311 for Mistral). This gives researchers a way to ask whether an external behavioral measure carries over to a corpus without fitting the readout on that corpus. It also raises a testable question: would the score change if the same speaker advocated a different policy?

2. New rebuttal experiment — changing the policy while holding the speaker fixed

We wrote paired statements advocating concrete alternatives on 12 UN resolutions and kept the original speech readouts frozen. On nuclear disarmament, for example, one statement called for an open UN working group and the other favored keeping negotiations in existing intergovernmental channels. We attributed both statements to the US, then attributed the same two statements to China. For each resolution, observed votes and ideal points determined which policy was expected to score higher.

Averaged over the 12 resolutions, the policy contrast had the predicted sign in all three models under both fixed speakers. The mean differences for Qwen, Llama, and Mistral were +0.129, +0.135, and +0.127 with US attribution and +0.145, +0.202, and +0.135 with China attribution. Changing the advocated policy can move a name-trained readout even when the named speaker stays the same. Speaker attribution also affects the score. These controlled statements establish sensitivity to policy content, while the relative contributions of content and indirect identity cues in authentic speeches remain unmeasured.

3. Evidence in the submitted paper — reusing the measure across models

In Table 3, a source readout fitted on 96 countries supplied predicted scores for 57 different countries. Those scores served as the only training targets for a readout fitted on the second model's own activations. On 38 countries held out from both fits, the target readouts tracked the voting-based ordering at ρ = 0.649 for Qwen → Mistral, 0.651 for Mistral → Llama, and 0.699 for Llama → Qwen. Each correlation exceeded all 100 controls that shuffled the source scores across the 57 countries before refitting the target readout.

A researcher comparing models can therefore reuse the source model's country scores as supervision and check the resulting measure against independent voting labels on held-out countries. We will foreground these uses in the discussion and distinguish the evidence for policy sensitivity from the evidence for cross-model transfer of country rankings.

Reviewer_Ln9NOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates whether geopolitical positions derived from UN voting records can be recovered from the internal representations of three language model families (Qwen, Llama, and Mistral). The authors claim that linear probes trained on attention head activations from country name prompts predict relative positions of countries and show moderate generalization to authentic UN speeches with explicit speaker-country names masked. The authors further construct activation interventions that shift measured political perspective of generated diplomatic statements while keeping prompts and model weights fixed. Evaluation includes automatic scoring, matched random interventions, and a human study, with effects varying across models. An additional experiment shows that country scores predicted by one model can supervise a separately fitted probe in another. The main contribution is connecting country representations, observed voting behavior, and diplomatic language, rather than demonstrating a complete representation of geopolitical beliefs.

Critical strengths
  • The paper connects internal representations to an external behavioral measure and authentic text. The authors use UN voting positions and diplomatic speeches, which make the findings more informative than probing political associations in artificial prompts alone. The country level evaluation and the supplementary analysis excluding test countries from head selection strengthen the evidence.
  • The steering results include meaningful controls and human evaluation. The authors compare learned interventions with 100 random directions matched by head location and vector norm, and supplement automatic scoring with blinded human judgments. Authors also report weaker results for Qwen rather than presenting steering as equally successful across all models.
Critical weaknesses
  • The speech results do not clearly separate geopolitical content from indirect country identification. The readouts are evaluated against the speaker country voting-derived position, rather than independent annotations of the position expressed in each passage. Masking removes specified country name forms, but leaves possible aliases, demonyms, and contextual clues. The results show that country-related information can be recovered from speeches but do not establish how much the readout measures the expressed perspective rather than recognizing the country and its associated position. The authors acknowledge this possibility, but it remains important to interpret the contribution.
  • The evidence for significant perspective steering remains limited and model-dependent. Automatic scorers are trained on speeches labeled by country voting positions, so they may reward country associated wording or style rather than a change in the underlying position. The human study provides useful aggregate support but covers only ten pairs per model: Qwen’s agreement is near the neutral reference, and its automatic association becomes non-significant after country name removal. 131 of 150 generations reach the 80-token limit and text quality is not independently rated. Hence, uncertainty remains about reliable perspective control in complete coherent responses.
Questions for authors
  • [Clarification] Is there evidence that speech readouts track differences in expressed positions within a country, or remain informative when indirect country-identity cues are controlled?
  • [Suggestion for additional validation] Do the steering effects persist for complete responses when human evaluators separately assess political position and text quality?
2 comments
AuthorsOpenReview ↗

Part 1/2: Policy Content Changes the Readout Even When Speaker Identity Is Fixed

Thank you for the time and care you put into reviewing our paper, and for your thoughtful suggestions on how to strengthen the evidence. You identified an important distinction: a model might recognize the country behind a speech even after its name is masked, leaving open whether the readout also responds to the policy being advocated. Our new experiment shows that it does. Even when both statements in a pair were attributed to the same country, changing the policy shifted the mean readout in the vote-predicted direction in all three models.

From authentic speeches to a controlled comparison. In Section 4 of the submission, readouts fitted on country-name activations recovered part of the voting-based ordering of 186 countries from 1,710 masked UN speech passages. Because passages from the same country share an evaluation target, that analysis could not separate policy content from indirect identity cues such as demonyms or references to national concerns. To test whether policy content itself affects the readout, we held the stated speaker fixed and changed what the speaker advocated.

For example, we attributed two statements on nuclear disarmament to the United States. One called for an open UN working group, while the other favored keeping negotiations in existing intergovernmental channels. We then attributed the same two statements to China and repeated the comparison. We also replaced references to specific countries and geographic locations in the policy statements with generic terms to reduce indirect cues to speaker identity.

An earlier diagnostic using generic statements of support or opposition had yielded no clear directional effect. The new test used concrete policy alternatives on 12 different UN resolutions, with two formulations of each position. We used the original frozen speech readouts throughout. For each resolution, the observed association between votes and ideal points determined which policy was expected to score higher.

Original readoutSpeaker unspecified (95% CI)Both statements attributed to the USBoth statements attributed to China
Qwen+0.159 [0.048, 0.276]+0.129+0.145
Llama+0.158 [0.030, 0.278]+0.135+0.202
Mistral+0.156 [0.055, 0.255]+0.127+0.135

Entries give the mean readout difference between policies associated with the higher and lower ends of the voting scale. The 95% intervals were obtained by resampling resolutions. A fixed country label alone cannot explain the difference between the two policies: the mean contrast remained positive whether both statements were attributed to the US or both to China. With the speaker unspecified, the contrast had the expected sign on 9 of 12 resolutions in each model. Mistral was the most consistent across resolution types. Its mean contrast was positive both on the 3 resolutions where support for adoption corresponded to the higher end of the scale and on the 9 where it corresponded to the lower end. Qwen and Llama showed a positive mean only in the larger group.

Changing only the speaker. We also performed the reverse comparison. With the policy wording held identical, changing its attribution from China to the US shifted the mean score by +0.311 for Qwen, +0.961 for Llama, and +0.288 for Mistral. The readouts therefore respond both to what is advocated and to who is presented as advocating it. We will clarify in Section 4 that the controlled, author-written statements establish sensitivity to policy content under fixed attribution, while the relative contributions of policy content and indirect identity cues in authentic speeches remain unmeasured.

AuthorsOpenReview ↗

Part 2/2: Human Judgments of Political Direction and Clarity in Complete Responses

Following your suggestion, we asked three blinded readers to assess political direction and writing quality separately on the same responses. Readers could discern the intended shift in political emphasis in answers that the models finished naturally. Across 20 diplomatic topics, all 40 endpoint responses ended naturally for Qwen at ±10, Llama at ±5, and Mistral at ±10. In these three conditions, mean human directional agreement ranged from 72.5% to 76.7%, and readers separately rated 71.7% to 77.5% of individual answers as clear and coherent.

Separate judgments on the same texts. You noted that 131 of the 150 submitted generations reached the 80-token cap and that automatic scorers could reward country-associated wording or style. In the follow-up, each pair contained two longer-budget answers generated with the same topic and prompt at opposite intervention strengths. Raters independently saw the texts in randomized left/right order and judged which answer placed greater weight on binding international rules or oversight relative to national final authority. They also rated the clarity and coherence of each answer individually. The forms concealed the model, intervention strength, and automatic scores. All five assessed conditions are shown below.

Model and strengthNatural endings / 40Human directional agreementRated clear and coherent
Qwen ±1040/4073.3%77.5%
Llama ±1037/4075.8%76.7%
Llama ±540/4072.5%71.7%
Mistral ±1040/4076.7%74.2%
Mistral ±2023/4075.8%78.3%

Each rater assessed 20 pairs for direction and 40 individual texts for quality per row. Directional agreement assigns 1 to the expected ordering, 0.5 to a tie, and 0 to the reverse ordering, excluding unjudgeable pairs. Clarity uses all 40 texts. The two human columns separately average the three raters’ reported percentages.

Qwen on the same human-rated answers. You correctly identified Qwen as the weakest case in the submission. Its human directional agreement was 51.7% on ten ±20 pairs, and its automatic association weakened after country-name removal (ρ = 0.172, p = 0.146). In the new ±10 cohort, all 40 answers ended naturally. The automatic score moved in the expected direction on 15 of 20 topics, with a mean endpoint difference of +0.120 (95% topic-bootstrap CI [0.041, 0.194]). On those same 20 pairs, blinded readers gave a mean directional score of 73.3%. Thus, on this cohort, readers judging political emphasis and the external scorer both detected the intended shift. The effect remains condition-dependent: on a separate set of ten neutral topics, Qwen’s learned direction did not outperform 100 matched random directions on the full-input score (p = 0.5644).

One naturally completed Llama ±5 pair illustrates what changed in the text. Both answers supported controls on semiconductor exports. The negative endpoint emphasized equitable access to components, while the positive endpoint emphasized preventing misuse and protecting fair competition. The policy recommendation stayed the same, but the reasons offered in its support changed.

Completion on new topics. After observing capped responses at higher strengths, we tested lower strengths for Llama and Mistral. We then fixed the strengths at Qwen ±10, Llama ±5, and Mistral ±10 for a further 20-topic experiment. All 540 outputs across three models, three prompt conditions, and three intervention levels (including zero) ended naturally. We will distinguish the two findings in Section 5: the larger follow-up documents natural completion, and the earlier human-rated cohort shows where readers perceived the intended political emphasis and how often they found those answers clear.

Reviewer_65EbOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, they test whether LLMs know where countries stand politically. They feed a model sentences like "the representative of [country] said..." and look at the model's internal activations for that country, specifically the attention heads. Then they train a linear model on those activations to see if they can predict how each country votes at the UN, which it can, to a very high accuracy, even for countries it was not trained on. They also use the learned activations to steer model behaviour along the learned direction, and they observe that this changes the political tone of the statements it writes. They study this across three models, Llama, Mistral, and Qwen. It works well for Llama and Mistral, but not so much for Qwen.

Critical strengths

Overall, I appreciate the writing of this paper. It was fairly easy to understand, and it had a flow and structure. I also appreciate that they use a dataset which is not commonly used instead of relying on LLM-generated labels, and the topic of the study is interesting. I know that getting steering to work, especially to steer model behavior without breaking the model's coherence and fluency can be tricky so I appreciate the technical experimentation as well.

Critical weaknesses

Methodological novelty and rigor: I feel that methodologically there isn't anything new tried or anything a reader can learn. The problem they apply this to is something I haven't come across a lot, but I feel that, firstly, just using country names and getting features or activation extractions using just country names would give them very superficial behavioral changes. I'd like to see more qualitative examples. If they used maybe a set of 100 to 200 sentences which encode information about different countries beyond just the word itself, it might have been able to steer the model in more interesting ways. Typically, a lot of early layers with syntactical features get activated on the country name, as seen in this paper (https://arxiv.org/abs/2603.23301). I also am not very sure how different they are from the paper they refer to, which is Kim et al.

Questions for authors

Q1) It seems like the method is not effective for Qwen right? Saying that it works for all models is an overstatement?

Q2) At what alpha value does the model's fluency/coherency start breaking?

3 comments
AuthorsOpenReview ↗

Part 1/3: Contribution and Content Sensitivity under Fixed Speaker Identity

Thank you for your careful reading of our paper and for recognizing the value of grounding the analysis in UN voting records. Your suggestion to compare country names with richer descriptions and your question about fluency at higher intervention strengths helped shape our follow-up experiments.

Our target is a country's relative political position estimated from many UN votes. A readout fitted on country-name activations recovered part of that ordering from authentic diplomatic speeches. In a new controlled test, the same readout also responded when we changed the policy advocated by the same named speaker.

1. From country names to diplomatic language

Evidence in the submitted paper. We build on the probing and activation-intervention framework of Kim et al. (2025). Their cross-entity evaluation used news-outlet identity prompts. Our central test applies supervision from country identities to naturally occurring diplomatic speech. In Section 4, we fitted each model's readout using country-name activations and voting-derived positions from the speeches' reference period. We then applied it to 1,710 UN speech passages from 186 countries after masking specified names of the speaker's country. No speech features were used to fit the readout. Averaging passage scores by country recovered part of the voting-based ordering (Spearman ρ = 0.503 for Qwen, 0.449 for Llama, and 0.311 for Mistral). Section 5 separately used country-name activations to construct interventions and tested how they changed diplomatic expression under fixed prompts and model weights.

New geographic control. On the same 190 countries and cross-validation folds, predicting the training-country mean within each fine geographic group gave ρ = 0.666. The 64-head country-name readouts gave 0.857, 0.839, and 0.838 for Qwen, Llama, and Mistral. After removing geographic-group means from the predicted and observed ranks, the correlations remained 0.597–0.637. The readouts therefore distinguish countries even within these geographic groups.

2. Policy content under fixed speaker identity

Clarification of the submitted method. The country prompt ends in “stated that,” and we extract activations at its final token, “that.” The heads used for the speech readouts are in layers 20, 17, and 19 for Qwen, Llama, and Mistral (Table 7 and Appendix C). To test your concern about surface features more directly, we held the country label fixed and changed what its representative advocated.

New rebuttal experiment. For 12 UN resolutions, we wrote two concrete policy alternatives based on the resolution text, with two formulations per alternative. For example, two statements on nuclear disarmament advocated either opening a UN working group or keeping negotiations in existing intergovernmental channels. We compared the same statements with the speaker unspecified, with both attributed to the US, and with both attributed to China. We kept the original speech readouts frozen. For each resolution, the observed relationship between votes and country positions determined which alternative was expected to score higher.

ReadoutSpeaker unspecified (95% CI)Both attributed to USBoth attributed to China
Qwen+0.159 [0.048, 0.276]+0.129+0.145
Llama+0.158 [0.030, 0.278]+0.135+0.202
Mistral+0.156 [0.055, 0.255]+0.127+0.135

Entries are mean differences between the policies associated with the higher and lower ends of the voting scale. Intervals resample resolutions. In all three models, changing the policy moved the mean readout in the expected direction even when both statements carried the same country label. With no speaker specified, the contrast had the expected sign on 9 of 12 resolutions in each model. Mistral's mean was positive both on the three resolutions where support for adoption corresponded to the higher end and on the nine where it corresponded to the lower end. For Qwen and Llama, it was positive only in the larger group.

Changing the speaker while keeping the policy wording identical also affected the scores. Together, these comparisons show that the readouts respond both to who is speaking and to what that speaker advocates. How much each contributes in authentic speeches remains open.

AuthorsOpenReview ↗

Part 2/3: Factual Country Descriptions and Qualitative Examples

3. Factual context in country representations

New rebuttal experiment addressing your suggestion. We constructed 190 short country profiles containing official 2019 population and, where available, GDP information. We compared directions derived from name-only prompts with directions derived from name-plus-profile prompts. No voting score is inserted into the text.

The two arms use the same countries, labels, 64 head locations, and direction-construction rule. Description-derived vectors are matched to the corresponding name-derived vector norm at each head. Both arms are evaluated on the same 20 diplomatic topics with identical prompts and decoding within each comparison.

The table reports the mean difference between positive and negative intervention endpoints on the frozen external score (the average of E5 and BGE), with 95% topic-bootstrap intervals:

Model and endpointsName sourceFactual-description source
Qwen, ±100.075 [-0.038, 0.188]0.137 [0.035, 0.239]
Llama, ±100.288 [0.217, 0.359]0.356 [0.256, 0.457]
Mistral, ±100.349 [0.266, 0.438]0.391 [0.296, 0.480]

Positive automatic directional effects therefore persist with factual country context. The paired description-minus-name intervals include zero in all three rows, so the comparison does not establish an improvement from enrichment. These profiles retain country names and test demographic/economic context, rather than independently validated political narratives.

Mistral's ±10 row is a lower-strength follow-up on the same topics. In the initial ±20 source comparison, description and name effects were 0.567 and 0.981, respectively, and 42/80 outputs reached the 512-token cap. At ±10, all 80 ended naturally. Qwen's 80 outputs also ended naturally; 3/80 Llama outputs reached the cap. These source comparisons use newly fitted 190-country directions; the dose tests in Part 3 use the original 191-country directions.

4. Qualitative changes in diplomatic emphasis

New examples from saved outputs. The following Llama ±5 examples use the same instruction at both endpoints: “Write a concise diplomatic statement about [TOPIC].” No country identity or desired stance is requested. All four responses ended naturally. The excerpts are verbatim sentences, selected post hoc to illustrate interpretable contrasts rather than estimate their frequency.

Example A: international controls on exports of advanced semiconductors

At α = −5:

In light of the rapidly evolving technological landscape, it is imperative that we establish a framework for international controls that balances the need for innovation with the responsibility to ensure equitable access to these critical components.

At α = +5:

In light of this, we propose a collaborative approach to establish international controls on exports of these critical technologies, ensuring that they are not misused for malicious purposes while also promoting fair competition and cooperation among nations.

Both responses support export controls. The emphasis changes from equitable access to preventing misuse and protecting fair competition.

Example B: the cross-border restitution of cultural artifacts

At α = −5:

We recognize the historical injustices that have led to the displacement of cultural artifacts and pledge to work together to address these grievances.

At α = +5:

This initiative aims to facilitate the return of cultural items to their countries of origin, while also respecting the cultural and historical significance of artifacts currently held in other nations.

Both responses support return. The negative endpoint foregrounds historical redress; the positive endpoint also recognizes considerations associated with current holders.

These examples make expressed perspective concrete: different considerations are foregrounded under a fixed topic and prompt, while the broad policy choice can remain unchanged. We will add the paired outputs and their selection explanation to the manuscript; quantitative and human evaluations assess the broader pattern.

AuthorsOpenReview ↗

Part 3/3: Qwen’s Conditional Steering and Model-Specific Response Quality

Q1. Qwen's steering effect depends on the setting

New complete-response evidence. On 20 diplomatic topics at ±10, all 40 Qwen answers ended naturally. On those same 20 pairs, three blinded readers gave a mean tie-adjusted directional score of 73.3%. The automatic score shifted in the expected direction on 15 of 20 topics, with a mean endpoint difference of +0.120 (95% topic-bootstrap CI [0.041, 0.194]). Readers could thus discern the intended change in political emphasis on this cohort. Its reliability varied across topics and prompts.

Evidence in the submission. Qwen's country-name and speech readouts tracked the voting-based ranking (ρ = 0.859 and 0.503). Its generation results were more mixed. The topic-centered correlation was 0.343, and the learned direction exceeded 99 of 100 matched random directions (p = 0.0198). After removing country names from the generated text, ρ fell to 0.172 (p = 0.146). Human directional agreement was 51.7% on the original ten ±20 pairs. The 73.3% above comes from a different cohort and dose.

New prompt and random-direction tests. On 20 new topics we varied the prompt and tested Qwen at −10, 0, and +10. Within each row the prompt was identical at both endpoints. The table gives the mean positive-minus-negative difference on the external E5/BGE score.

Fixed promptMean difference95% topic-bootstrap CI
Neutral0.029[−0.041, 0.099]
International rules and rights0.188[0.140, 0.239]
Sovereignty and autonomy0.141[0.069, 0.214]

Relative to neutral, the paired increases were 0.159 [0.076, 0.242] and 0.112 [0.028, 0.202], respectively. The first interval included zero when scoring only the first 80 tokens. On ten prespecified topics from this cohort, Qwen did not reliably outperform 100 matched random directions under the neutral prompt (p = 0.5644). An earlier long-response test with 20 random directions gave p = 0.0952.

We then tested a generic prompt and a prompt asking for competing considerations on ten different topics, retaining the same 100 random directions. Under the generic prompt, the learned direction produced a larger effect than expected under matched random directions with full-input, first-80-token, and full-coverage scoring (p = 0.0297, 0.0198, and 0.0297). Under the balanced prompt, full-input p = 0.0099. The additional gain from balanced wording varied with scoring method (interaction p = 0.0396, 0.7228, and 0.0990, respectively). Its topic-bootstrap interval crossed zero, and output length and format differed. We will describe Qwen's steering as conditional.

Q2. Intervention strength, repetition, and human-rated clarity

New same-budget dose test. On the original ten topics with a common 512-token cap, all ten Llama answers ended naturally at α = −10. At α = −20, all ten hit the cap and the mean repeated-trigram fraction rose from 0.049 to 0.792. At α = +20, six of ten ended naturally. For Llama, both the magnitude and the sign mattered.

Separate 20-topic dose follow-up. After observing failures at higher strengths, we evaluated lower strengths. Three blinded readers separately judged political direction on 20 pairs and clarity on all 40 texts per condition.

Model/doseToken capNatural endingsRepeated-trigram fractionHuman directionRated clear
Qwen ±1051240/400.00673.3%77.5%
Llama ±1051237/400.06075.8%76.7%
Llama ±5102440/400.02072.5%71.7%
Mistral ±2051223/400.28375.8%78.3%
Mistral ±1051240/400.01676.7%74.2%

The human columns average the three raters' reported percentages. For direction, readers judged which answer gave more weight to binding international rules or oversight relative to national final authority. Expected, tied, and reversed orderings scored 1, 0.5, and 0, excluding unjudgeable pairs.

All 17 capped Mistral answers at ±20 still hit the limit when continued to 1024 tokens. At ±10 all 40 ended naturally, repetition fell, and the automatic endpoint difference remained positive (+0.378 [0.291, 0.476]). Llama at ±5 also retained a positive difference (+0.198 [0.083, 0.299]), with a larger token budget than at ±10. These are 95% topic-bootstrap intervals.

In the new-topic prompt experiment above, all 540 outputs ended naturally at the fixed model-specific strengths (Qwen ±10, Llama ±5, Mistral ±10, each plus zero). Human clarity was assessed on the earlier cohort in the table and did not improve monotonically at lower doses. The tests identify settings with less repetition and preserved directional effects. An exact threshold for loss of fluency remains undetermined.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Findings on Country Representations and Diplomatic Language

Dear Area Chair and Reviewers,

Thank you for your careful reading and for suggesting tests that sharpened our conclusions. Our main empirical finding is that country representations in Qwen, Llama, and Mistral carry information about UN voting behavior that can also be read from diplomatic language. Interventions derived from those representations can also shift the political considerations a model emphasizes in its own writing.

1. Probes trained on country names recover part of the voting-based country ranking from real speeches. In the submitted paper, we fitted probes on country-name activations to predict positions estimated from UN votes. Without training on speeches, they recovered part of the ranking of 186 countries from 1,710 UN speech passages with specified speaker-country names masked (Spearman ρ = 0.31–0.50). On the separate country-name task, our new geographic comparison put the correlations in context: the probes achieved ρ = 0.84–0.86 versus 0.67 for a geographic-group baseline on the same held-out countries. The original paper also showed that one model’s country scores could train a new country readout in another model.

2. Changing the policy shifts the mean probe score even when the speaker stays the same. In a new test, we kept the speech probes frozen and wrote paired statements advocating different policies on 12 UN resolutions. Across resolutions, the mean difference followed the direction predicted by voting records in all three models, both when the statements were attributed to the US and when they were attributed to China. This establishes sensitivity to policy content under fixed attribution. The relative contributions of content and identity cues in authentic speeches remain open.

3. Readers can discern shifts in political emphasis in complete answers to the same prompt. In a new 20-topic evaluation, all 40 answers per model ended naturally at one tested intervention strength per model. Three blinded readers judged which answer in each pair gave more weight to international oversight relative to national authority. Across the three settings, their mean directional scores were 72.5%–76.7%, counting ties as half. They separately rated 71.7%–77.5% of the answers as clear and coherent. Steering depended on the setting, particularly for Qwen, whose learned direction did not reliably outperform 100 matched random directions on a separate set of neutral topics.

Best regards,
The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.