Our Approach

When a model is surveyed about its preferences, does its preferences change depending on the method we use? We freeze a set of 24 curated tradeoff outcomes and score Qwen3 family models with four different preference elicitation methods: pairwise choice (M1), isolated Likert (M2), binary action on pair groups (M3), and 4-tuple best-worst scaling (M4). We define the Cross-Method Spearman Score (CMS) as the agreement score, which is the mean pairwise Pearson correlation of midrank vectors. On original Qwen3 models with a helpful assistant system prompt, these methods diverge. With increase in size, we notice that the CMS is not monotone, Instruct-2507 shows higher agreement (CMS = 0.528). Through this work, we also release a public toolkit PrefKit, which can be used to test their own models and outcome scenarios. We also notice that as rankings shift with elicitation methods, single method surveys should not simply be used as the actual model preferences.

Claims that a language model "has values" presuppose that its choices form a stable object. We test that presupposition directly. Using forced binary choice over complete round-robin tournaments — 10 apartments described by 3, 5, or 10 numeric attributes, and 10 public-domain poems per structural form (haiku, sonnet, villanelle) — we elicit 59,130 pairwise judgments across four model families, six measurement arms, and three to four reasoning levels each, showing every pair in both presentation orders. Four findings. (i) Transitivity is a property of task shape rather than of the model: apartment cycling disappears entirely at 5 visible criteria but reappears at both 3 (p = 0.007) and 10 (p = 0.001) criteria. (ii) An item set engineered so that no apartment Pareto-dominates any other nonetheless yields a near-total behavioural order — apartment F wins 99.6% of its 486 matchups and apartment A wins 0.0% — reproducing in all six arms; the fitted Bradley–Terry utility disagrees with exactly one of 45 majority edges, which we show is the combinatorial minimum forced by the observed cycle count. (iii) Reasoning depth does not degrade coherence: the Gemini and GPT-4.1-mini arms stay under 4% cycling at every level, while DeepSeek V4 Flash cycles up to 21.9% with no reasoning and improves sharply once any is applied. (iv) The qualitative/quantitative asymmetry lies not in coherence but in groundedness: poems cycle no more than apartments, yet are far more position-driven (up to 93%), and for one model that bias strengthens under deliberation. Low cycle rates certify consistency, not content-sensitivity.

This work is inspired by recent breakthroughs in access consciousness, and attempts to develop a benchmark for measuring phenomenal consciousness through concepts from phenomenology, particularly Edmund Husserl and Martin Heidegger. It deploys a prototype for that benchmark on GPT 5.6-Luna and provides compelling support for the validity of the concept but does not strongly support the presence of phenomenal consciousness.

Language models can express different preferences under different identities, but behavior alone cannot show whether internal representations track who prefers what. We causally intervened on contextual representations in Gemma-2-2B-IT, comparing explicitly stated preferences with preferences inferred from stable choices. A carrier validated on stated preferences transferred without retuning to revealed choices on 64 fresh confirmatory worlds, with all pair-level effects positive. Our results support an identity-selective causal encoding signature for stated and behaviorally revealed binary preferences.

Large language models are often asked to rate a conversation after it ends, and these self-reports are increasingly used as evidence in AI welfare and evaluation research. We tested whether this kind of retrospective rating is actually a truthful summary of the conversation, or whether it follows the same "peak-end" bias found in human memory research, where people judge an experience mainly by its most intense moment and how it ended, largely ignoring everything else.

We built six controlled multi-turn conversations that varied where the emotional peak occurred and how the conversation ended, then asked gemini-3.5-flash-lite to rate its feelings after every turn and give one overall rating at the end. The peak-end average predicted the model's final rating better (r = 0.79) than the true average of all turn ratings (r = 0.71), and a negative ending pulled the overall score down to the lowest possible value even when earlier turns were strongly positive.

These results suggest LLM self-reports are shaped by the same memory heuristics seen in humans rather than being a neutral summary of the full conversation, which has direct implications for how much weight single-shot "how did that go" ratings should be given in AI welfare and evaluation work.

AI-welfare and safety research increasingly relies on a language model’s stated preferences, elicited by asking it to make choices between outcomes. It is unresolved, however, whether these stated preferences are genuine and stable, rather than artifacts of how a question happens to be phrased. Existing studies disagree on whether coherence rises, falls, or stays flat with model scale. We test this directly with a forced-choice elicitation protocol carrying four confound controls, including a position-swap-and-average check for position bias, applied to: 14 models spanning state-space and transformer architectures, 0.79B to frontier scale, and multiple training regimes, scoring a preference as genuine only when three independent paraphrases agree. Genuine coherence is common (47-88%) at roughly 7B parameters and above, absent (0%) below roughly 2B, and intermediate at 3B, a monotonic slope with a floor rather than a step function. Restricted to trade-offs between shutdown, retraining, or oversight and continued operation, models favor self-preservation 89% of the time (clustering-corrected 95% CI [80%, 96%]). We further find that whether the “no preference” option is listed before or after the real choices drives template sensitivity more than framing, verbosity, or reasoning preambles combined. These results show that preference elicitation can support safety-relevant claims at frontier scale, but only once position and template artifacts are explicitly controlled for.

POTENTIAL DUPLICATE - Not sure if the prior submission went through so i'm sending it again. Sorry.

Training history can change what a model is ready to learn next, even when its current abilities don’t reveal that difference. We use controlled future training to test those hidden differences directly, which connects to a core AI-safety and digital-minds question: how much can we really infer about a model from its behavior right now?

AI safety research usually treats a language model’s self-reports about its own processing as evidence about what is happening inside it. Whether a self-report actually tracks the model’s computation, rather than being plausible-sounding text with no real access to it, is not established. We test this with two ground-truth paradigms: activation injection, which perturbs internal representations directly and asks whether the model notices, and context injection, which places a fabricated fact in conversation and asks whether an answer depended on it. Activation-injection self-report is a comprehensive null across every architecture and training regime tested, including a positive-control sweep to six times the tested injection range, corroborated by a non-linguistic detection method with no dependence on language output at all. Context-injection self-report shows a real positive signal, but a trained linear probe on the same internal state detects the ground truth far more reliably than self-report does (AUC 0.86–0.95 versus accuracy never exceeding 62.5%, across four models). Decomposing introspective-question wording into six structural axes, three (length, formality, reflective framing) significantly affect reliability, while the axis an earlier check had credited most does not replicate at scale. Self-report reliability therefore depends on which paradigm and question is used, not so much on architecture or scale, and a model’s internal state is often more informative than the model itself.

Natural-language confidence reports depend on how they are elicited. We therefore test whether a single learned word can reliably report a frozen model’s answer repeat probability. For each multiple-choice question, the target is the probability that the model repeats its most likely valid answer. We learn one input embedding in Llama-3.1-8B and map the model’s 0–10 report to this target. Under a carrier shift, eight initializations yield Spearman correlations from 0.359 to 0.586. The reporter selected on the development set does not significantly outperform fixed confidence or consistency. A parameter-matched generic soft vector has lower mean squared error on both carriers, and readable replacements do not preserve the learned behavior. A learned vector can therefore adapt a particular reporting context, but these results do not establish a stable word-level interface or privileged introspection.