Aug 16, 2026

Behavioural Stability and Context Sensitivity in Large Language Mode

Shruti Tripathi

This project investigates how reliably we can measure apparent preferences in AI models. We evaluate GPT-5-mini and Gemini 2.5 Flash across 8 preference dimensions, 4 elicitation methods, and 4 contextual conditions, with repeated trials.

Rather than treating model self-reports as direct evidence of internal experiences, we measure behavioral preference stability, agreement across elicitation methods, and sensitivity to contextual framing. We use paired statistical tests and bootstrap confidence intervals to compare models and identify when measured preferences change under different experimental conditions.

The goal is to develop a more rigorous methodology for studying AI preferences and welfare-related signals while avoiding unsupported claims about consciousness or subjective experience.

Review Project

See Code

View Related Sprint

Reviewer's Comments

The current project focuses on a very clear and specific methodology-based question: How stable are apparent AI preferences when either the preference measurement (elicitation) or framing of the preference changes? In order to assess this question, the authors ran a total of 512 different experiments assessing GPT-5-mini and Gemini 2.5 Flash across eight separate preference dimensions. They assessed these using four distinct types of elicitation mechanisms (methods) and four different contextual conditions.

This study has several notable strengths. A first major strength is that the study is very cautious in terms of exactly what it does and does not claim its results prove. As opposed to interpreting model-generated reports of self-preferences as evidence for either the existence of consciousness or welfare within an artificial intelligence system, the study interprets such reports as simply another type of measurable behavior. Thus, the study asked whether the measurable behaviors reported by the models are actually reliable.

Therefore, the study's use of repeated trials, paired comparisons, bootstrap confidence intervals, and FDR correction provides a strong statistical framework for analyzing the data collected. Moreover, the fact that Gemini demonstrated greater repeated-trial stability than GPT-5-mini, while both models demonstrated significant context sensitivity, provides important insight into the fact that even if apparent AI preferences are reliably measured, they need not be completely independent of the context in which they were generated.

One major weakness of this study is the relatively small number of models tested, the low number of repetitions for each experimental treatment cell, and the restricted set of prompt formulations used to generate the preference trade-off pairs. Although this made it possible to report some interesting findings regarding preference stability among the models tested, it limits how broadly those results can be applied. Furthermore, although the deterministic generation settings reduced sampling variability by minimizing randomness during generation, they do not capture how preference stability would behave under more typical and realistic stochastic usage patterns for these models.

In order to strengthen this project further, additional model families could have been assessed, more repeated trials per treatment condition could have been run, and a much larger variety of prompts could have been developed to elicit preferences. Additionally, it would be beneficial to determine whether the behavioral stability observed at the output level correlates with any persistent aspects of the models themselves.

The project investigates the stability and context sensitivity of preferences in two large language models, GPT-5-mini and Gemini 2.5 Flash, through a controlled experiment with multiple preference trade-offs, elicitation methods, and contextual conditions. The study demonstrates that Gemini 2.5 Flash shows higher repeated-trial stability compared to GPT-5-mini and highlights significant reversals in preferences influenced by contextual framing, particularly with persona-autonomy framing. These findings are valuable for understanding the variability in model behavior across different experimental setups.

However, the methodological design has some limitations. The study only conducts a couple of repeated trials per experimental cell, which limits the robustness of the stability measure. Additionally, the experiment focuses on only two models and eight preference trade-offs, which restricts generalizability to other language models or contexts. The analysis also does not account for potential artifacts from prompt phrasing or the influence of assistant behaviors, which could affect the observed outcomes.

To improve the study, increasing the number of repeated trials would provide a more reliable measure of stability. Expanding the experiment to include more model families and various sizes, as well as different prompt rewordings, would enhance generalizability. Additionally, incorporating activation-based or causal techniques alongside behavioral preference assessment could help determine if stable preferences reflect lasting internal frameworks or context-sensitive language generation.

Cite this work

@misc {

title={

(HckPrj) Behavioural Stability and Context Sensitivity in Large Language Mode

author={

Shruti Tripathi

date={

8/16/26

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