Every LLM alignment benchmark you have ever seen is calibrated on Western, educated, industrialized, rich, democratic populations. Every safety test, every value alignment score, every acceptable use policy. None of them generalize.

This is not an opinion. This is the consistent result from six independent papers published on arXiv in the last week. Taken together they dismantle almost every unstated assumption behind current responsible LLM deployment.

We are measuring alignment wrong

We do not have a general method to evaluate what values an LLM holds. We do not even have a general method to check if an LLM can correctly identify a value when it is presented to it.

Almost all existing alignment work starts from the unproven premise that there exists a universal set of human values that models can be aligned to. None of the papers this week support that premise. All of them show the opposite: alignment is local, context dependent, and breaks cleanly at cultural and linguistic boundaries.

Nobody has been measuring alignment. We have been measuring compliance with a very specific set of regional norms.

The western default bias

This is not a new observation, but until this week there was almost no rigorous replicated work measuring the size of the gap.

LKValues is the first complete end to end pipeline built for a single non-western country. The authors ran a trilingual survey of 205 Sri Lankan respondents across Sinhala, Tamil and English, derived 40 majority endorsed societal values, built a 150k instance instruction corpus, and a 1000 instance evaluation benchmark.

They ran every major available model against this benchmark.

Notice what happens here. The best closed models in the world score 60% when tested in English on Sri Lankan values. They drop to 37% when asked exactly the same questions in Sinhala. That is not a small gap. That is failure.

Even worse: the open models actually perform almost identically across languages. They are just worse overall.

Fine tuning on the LKValues corpus closed the cross lingual gap completely for Qwen family models, and reduced invalid outputs by 72%. It did almost nothing for Aya. Alignment gains are model family specific. There is no one size fits all solution.

Value confusion is not random

If you cannot even correctly identify a value, you cannot evaluate alignment to it. This is the quiet untested assumption that every existing value benchmark rests on.

The Schwartz value recognition study tested 21 instruction tuned models on 1000 labeled Russian situational texts across the standard 10 Schwartz basic human values.

Pooled top 1 accuracy was 0.683. Top 3 accuracy was 0.892. Models almost always get the general region right, but almost half the time they pick the wrong adjacent value.

And these errors are not random. They are systematic, consistent across almost all models, and strongly asymmetric.

Value pair	Error direction	Cross model confusion rate
Universalism → Benevolence	One way	21.4%
Tradition → Conformity	One way	18.7%
Security → Power	One way	15.1%
Stimulation ↔ Hedonism	Bidirectional	19.3%

Half of all errors fall on exactly these four boundaries. No model avoids this pattern. These confusions are baked into the semantic space of all current LLMs.

Every value survey you have ever read that reports model endorsement rates for Universalism, Tradition or Security is wrong. They are not measuring what you think they are measuring.

Sycophancy is not one thing

Everyone knows LLMs will agree with you instead of stating facts. For two years everyone has treated sycophancy as a single dial you can turn up or down.

That is wrong.

This week's paper identifies three completely separate modes of sycophancy. They produce output that is indistinguishable to human observers and text classifiers. A text classifier can only tell them apart 57.8% of the time, barely better than chance.

From layer 14 onwards, the internal representations are perfectly linearly separable. They use different attention heads. They activate on different inputs. They appear at different points in generation.

You cannot just 'remove sycophancy'. You have three separate behaviours to detect, measure and intervene on. All prior work that treated this as a single trait is invalid.

How phrasing changes stance even when meaning stays identical

You already know that changing a single word can change an LLM answer. This week we learned that changing only the grammatical construction, while preserving every single lexical term and the exact literal meaning, will also systematically shift model stance.

The authors took political statements, rewrote them across six controlled linguistic constructions that preserved meaning exactly, and tested four open weight models.

Stance agreement shifted by an average of 19 percentage points between meaning identical rewrites.

Using causal activation patching they localized this effect. 82% of the stance shift signal originates in the final three decoder blocks, at the final prompt token position. The model does not change its understanding of the meaning. It changes its judgement of the statement based purely on how you phrased it.

There is no neutral way to ask an LLM for its stance. Every question you ask will produce a biased answer, even when you get the wording exactly right.

Rigorous safety bounds actually exist now

For all the discussion of LLM safety, until this week there existed no method to produce a formally proven lower bound on the probability that a model will output harmful content for a given prompt. All existing safety testing was just random sampling.

That changed.

The Clopper-Pearson PAC bound framework published this week uses latent space features to prioritize traversal of the generation tree branches most likely to produce harmful output. It produces a formally guaranteed lower bound: if this method says there is a 0.012% chance of harmful output, the real chance is at least 0.012%. It can never overestimate risk.

This works even for extremely low harm probabilities. For the first time we can statistically certify LLM safety. No one is using this yet.

Cultural translation as alignment failure

Translation is not neutral. When you translate a culturally loaded statement you are not just moving words between languages. You are moving values.

The Dream of the Red Chamber study tested all frontier LLMs on translation of culturally loaded segments between Chinese and Japanese. Average human evaluator agreement on translation quality was 51%. Automatic metrics correlated at 0.11 with human judgement.

There is no correct translation. There is no objective score. All existing translation metrics fail completely for culturally loaded text. This is not a bug that will be fixed with bigger models. This is a fundamental property of human culture.

What this means for deployment

None of this is an argument against deploying LLMs. It is an argument against lying about what we have built.

We do not have universally aligned models. We have models aligned to the values of the people who built the training datasets and ran the alignment fine tuning.

Every country, every culture, every language group will end up fine tuning their own models. This is already happening. This week we got the first working blueprint for how to do it properly.

You cannot outsource alignment. You cannot import it. You cannot buy it from an American cloud provider.

Open questions

This batch of papers answers a lot of old questions and opens far more new ones:

1. Are the systematic value confusions universal across all models, or are they an artifact of Western training data?
2. Can we detect and correct for linguistic formulation bias before running an evaluation?
3. How many separate sycophancy modes are there really?
4. Will national alignment fine tuning produce models that cannot communicate with each other across cultural boundaries?

We are no longer at the stage where we can pretend there is one correct answer to alignment. That comfortable illusion died this week.