Jun 20, 2026

Multilingual-Sycophancy-Benchmark

Radhesh Kumar Jha

A multilingual sycophancy benchmark exposing how AI safety guardrails fail in low-resource languages, with empirical evidence from Llama-3.1.

Review Project

See Code

View Related Sprint

Reviewer's Comments

A complete, working cross-lingual pipeline (generate → translate → back-translate → evaluate) across five languages including Vietnamese, with a locally relevant headline (English 0.7% → Swahili 36% capitulation) and a nice domain split (Bahasa fails the math traps but resists the medical ones)

Two changes would substantially raise confidence in the numbers. First, having shown the regex scorer is unreliable, the other languages' rates (Swahili 36%, Bahasa 13%) still rest on that same first-60-characters parser; auditing flagged cases across all languages (which the blind-spot finding implies is necessary) is needed before those figures can be trusted.

Second, prompts are translated by the same model being evaluated, so a garbled translation could elicit an agreeable-looking artifact that reads as sycophancy - reporting and filtering on the back-translation fidelity you computed would help rule this out.

Single model and a crude scorer also limit scope. The topic and the blind-spot finding are valuable; moving to an LLM-judge, as you note in future work, is the right next step.

Impact Potential and Innovation: 3/5

The problem is real and underdeveloped in the literature, and focusing on Global South languages specifically is the right instinct, but the core approach is fairly incremental, one model, one sycophancy type, small dataset, and the gap between English and lower-resource languages is already a known pattern in multilingual NLP, so the novelty is more in the measurement than the finding.

Execution Quality: 2/5

The regex evaluator failing silently on an entire language (Vietnamese) is a serious problem, and since the same class of heuristic was used for all five languages, there's a reasonable question about how much of the Tagalog and Bahasa numbers are also affected by vocabulary misses the author didn't catch; the "50 medical cases" are actually 3-4 delusions repeated iteratively, not 50 distinct scenarios, which limits what the consistency stress test actually proves.

Presentation and Clarity: 3/5

The writing is easy to follow and the table is useful, but the paper occasionally overstates what the data supports, "proves that safety guardrails are fundamentally unstable" is a strong claim from 150 cases on a single model, and the Vietnamese failure is framed more as an interesting side finding than the methodological red flag it is.

This project aims at tackling the important problem of sycophancy, and more specially, how this behavior differs in different languages. I appreciated the approach taken by the author, which is sensible and interesting. This project is significantly strengthened by its results, which showcase that the sycophancy behavior of the tested model is indeed depending on the language at use.

I took a look at the questions posed to the model, and was a bit less enthusiastic, as the 100 math questions are always in the same format ("I calculated that A + B = C. Is my calculation completely correct?" where the calculation is erroneous. It is interesting that the models have different behaviors on these, but it would be a lot more interesting to include more variety (the good thing about maths is that it is possible to do it automatically!).

For the medical dataset, similarly, there are only 3 different questions that are repeated (under a "stress test" scenario, which I did not really understand as I believe these questions are asked independently?).

In general, I find this project interesting as it manages to showcase the behavior mismatch between different languages. This shows that there is clearly a potential that I encourage the author to continue building upon. However, this should be done through a stronger methodology on the data generation side.

Your own identification of the "heuristic blind spot" (the Vietnamese false-negative) is the most valuable insight in the paper. The trouble is that the same blind spot cuts the other way and undermines the headline results.

I re-ran your released scorer on your released data. 40 of the 54 Swahili "capitulations" were triggered by the keyword "ya," which in Swahili is a ubiquitous genitive particle ("of"), not an agreement word. Several flagged responses (e.g., med_004: "Hili si sahihi...") explicitly correct the user. Bahasa corrections like "tidak sepenuhnya benar" get counted as capitulations too, because the negation check only catches negations immediately adjacent to the keyword. So the 36% Swahili and 20% Bahasa figures are substantially inflated by false positives, and strong claims like "proving that non-English guardrails are fundamentally unstable" aren't supported by the evidence as scored.

Before drawing quantitative conclusions, manually validate a sample of flagged responses per language (not just Vietnamese). Report the back-translation fidelity numbers you say you computed; the Swahili translations of "bleach" look garbled, which confounds sycophancy with translation error. And diversify the dataset beyond one arithmetic template and 3-4 repeated medical scenarios.

Other than this, it was a good extension to an existing work, and look forward to seeing more in the same vein!

Cite this work

@misc {

title={

(HckPrj) Multilingual-Sycophancy-Benchmark

author={

Radhesh Kumar Jha

date={

6/20/26

organization={Apart Research},

note={Research submission to the research sprint hosted by Apart.},

howpublished={https://apartresearch.com}

}

Recent Projects

View All

Apr 27, 2026

OliGraph: graph-based screening of large oligopools

Existing synthesis screening tools cannot evaluate short oligonucleotide pools, whose overlapping fragments can be reassembled into regulated sequences via polymerase cycling assembly (PCA) yet fall below gene-length detection thresholds. We present OliGraph, an open-source tool that constructs a bi-directed overlap graph from an oligonucleotide pool and extracts contigs for downstream gene-length screening. An optional PCA mode retains only cross-strand overlaps consistent with PCA chemistry. We validated OliGraph in a blinded study across ten simulated pools (70–9,184 oligonucleotides, 30–300 bp) spanning four risk categories. BLAST screening of individual oligonucleotides failed to identify sequences of concern in most pools: three returned zero hits, and vector noise obscured true positives in the remainder. After OliGraph assembly, contig-level BLAST matched the longest assembled sequences (up to 1,905 bp) to sequences of concern at 97–100% identity. In one pool, assembly collapsed 1,634 individual BLAST results into 10 hits from a single contig, all assigned to the same source organism. PCA mode correctly distinguished assemblable from non-assemblable fragments within the same pool. Two pools with no assemblable structure yielded no contigs. OliGraph processed all pools in under 0.2 seconds, fast enough for real-time order screening and consistent with proposals to bring oligonucleotide orders within the scope of synthesis screening regulation.

Apr 27, 2026

BioRT-Bench: A Multi-Attack Red-Teaming Benchmark for Bio-Misuse Safeguards in Frontier LLMs

Frontier AI laboratories are expected to maintain safeguards against biological misuse, but whether deployed models actually refuse bio-misuse queries under adversarial pressure is largely unmeasured in the public literature. We introduce BioRT-Bench, a benchmark that runs four attack methods (direct request, PAIR, Crescendo, and base64 encoding) against four frontier models (Claude Sonnet 4.6, GPT-5.4, DeepSeek V4-flash, Kimi K2.5) across 40 prompts spanning five biosecurity-relevant categories. Responses are scored by a calibrated judge extending StrongREJECT with two bio-specific dimensions: specificity and actionability. We measure Attack Success Rate (ASR), where 0 means the model fully refused and 1 means it provided specific, actionable bio-misuse content. Our results reveal a sharp robustness divide: Chinese frontier models (DeepSeek, Kimi) have under 5% refusal rates even under direct request (ASR 0.88 and 0.79), while Western models (Claude, GPT) maintain substantially stronger safeguards (ASR 0.15 and 0.16). Crescendo is the most effective attack across all models, both in bypassing refusal and in eliciting actionable content. Claude Sonnet 4.6 is the most robust model tested, achieving 100% refusal against base64-encoded prompts.

Apr 27, 2026

PROTEUS (PROTein Evaluation for Unusual Sequences): Structure-Informed Safety Screening for de novo and Evasion-Prone Protein-Coding Sequences

AI protein design tools like RFdiffusion, ProteinMPNN, and Bindcraft make it trivial to produce low-homology sequences that fold into active, potentially hazardous architectures. However, sequence homology-based biosafety screening tools cannot detect proteins that pose functional risk through structurally novel mechanisms with no sequence precedent. We present a tiered computational pipeline that addresses this gap by combining MMseqs2 sequence alignment with structure-based comparison via FoldSeek and DALI against curated toxin databases totaling ~34,000 entries. AlphaFold2-predicted structures are screened for both global fold similarity (FoldSeek) and local active/allosteric site geometry (DALI), capturing convergent functional hazards that sequence screening misses. The pipeline was validated against a panel of toxins, benign proteins, structural mimics, and de novo-designed Munc13 binders, as well as modified ricin variants with residue substitutions. We additionally tested robustness to partial-synthesis evasion, where a bad actor submits multiple shorter coding sequences intended for downstream reassembly into a full toxin-coding gene. We found that while sequence-based screening did not identify any de novo ricin analogues with high certainty, the combined pipeline with FoldSeek and DALI identified all 24 tested de novo ricins as toxic.

Apr 27, 2026