Continual knowledge-updating methods are often declared superior from one final checkpoint and one conventional adapter rank. We show that this can be insufficient to identify the better operating point. Holding a periodic hierarchy fixed, we compare it with cumulative replay over a 24-month Wikidata stream while varying evaluation month, replay LoRA rank, and query formulation. The apparent winner changes across this region: on Qwen2.5-1.5B, the hierarchy's 5.0-point advantage over rank-8 replay becomes an 11.6-point deficit against rank-72 replay, and at high ranks a consolidation-aligned endpoint can suggest a tie while time-averaged replay leads by 9-13 points. The same rank-conditioned reversal appears on Llama-3.2-1B and held-out paraphrases. These results show that method ranking in continual updating can depend jointly on when performance is measured and how much replay-side adaptation capacity the baseline receives. We therefore propose reporting trajectories and capacity sweeps, and declaring a robust winner only when the ordering is stable across the evaluation region; otherwise, comparisons should report winner regions and retention-stability-cost frontiers. Under this protocol, the periodic hierarchy is a lower-update-cost operating point, not a quality winner.
Activation steering is often evaluated under the answer encoding used to construct the direction. A reported gain may reflect the intended judgment or compatibility with answer identifiers seen during construction. We introduce Cross-Encoding Steering Evaluation, which freezes an intervention while re-encoding answers to the same held-out items. On NormBank, after A/B/C identifiers are reassigned, contrastive activation addition (CAA) induces larger target-versus-source score changes for the extraction indices than for the semantic labels under the new mapping. We call this extraction-index following. Varying identifier vocabulary (A/B/C, X/Y/Z, or 1/2/3) and row order shows that the effect tracks extraction index rather than row position. After matching direction norms across layers, extraction-index following emerges mainly at later depths. A low-rank output-sensitive component containing 15.4% of the direction's squared norm retains 96.3% of this effect. An Inference-Time Intervention (ITI)-style method also favors extraction-index over semantic-label following on NormBank in three models. In aggregate, MNLI favors extraction-index following, whereas Social Chemistry 101 (SC101) favors semantic-label following. Multiple-choice and open-ended evaluations can yield different behavioral conclusions. Thus, a steering gain under one answer encoding does not by itself identify what the intervention controls.
Whether a language model has improved itself is increasingly judged not by mean accuracy but by which individual problems it gains and loses. Tracking these transitions means differencing two noisy estimates, leaving them vulnerable to measurement artifacts. Auditing three rounds of rank-$32$ LoRA self-training on Qwen3-8B against a frozen control pushed through the identical pipeline, we identify seven measurement failures, each of which inverts a reported finding when its control is absent. Several are standard practice. A ledger built on a single greedy decode manufactures capability changes on an untrained model, largely an artifact of inference batching; the expansion statistic separating acquisition from sharpening assigns that same model a rate of $0.280$. The natural threshold repair does not survive replication: estimated across the frozen comparisons such a design already contains, its null stays non-zero. We replace it with a per-problem exact test against a pooled baseline under false-discovery-rate control, which detects nothing on any held-out replicate and is unchanged under the multiple-testing rule, error rate and pool size. Applied to a ladder of arms matched in stream, volume and evaluation, the audit finds that external distillation improves problems the base model rarely reaches while three forms of self-training do not; a regression rejects this asymmetry as a by-product of distillation's larger overall gain ($p < 10^{-8}$). On the far smaller set of problems the base model never reaches, the evidence is inconclusive, while self-training corrupts problems solved at baseline at rates well above the measured floor. Transition-level auditing therefore requires a separately measured null for every statistic it reports: nulls that cost no new experiments, built from baseline replicates a multi-arm study already owns, though not from as few as most possess.
Take three frontier mixture-of-experts models (Alibaba, OpenAI, NVIDIA; 3.6-4.0B active parameters each) and fine-tune them to reason in a low-resource language. On accuracy benchmarks almost nothing happens, and the benchmark itself is noise at this scale: changing only the random seed moves the score by 7.7 points, more than every data and recipe effect we measured. That null is our first result. The real changes live where accuracy cannot see. Base models never think in Greek: 0 of 1,000 reasoning traces, even when the question is Greek, so the model answers correctly while reasoning in a form its user cannot read, audit, or correct. After supervised fine-tuning (SFT), every released checkpoint reasons in the language of the question on ~98% of items, one family at 3x fewer tokens, with judged grammaticality improving on all four models and general ability within a few points of each base: nothing was forgotten, and fluency was gained. We propose six behavioural dimensions that make such changes measurable, each gated to reject any metric that correlates with output length, and we report how our own instruments lied: six failures, each caught by a control. What SFT cannot do is fix its own defects: a quarter of answers skip the requested format, answers leak into the reasoning channel, and an explicit "think in English" is obeyed under half the time. Reinforcement learning with verifiable rewards, pre-registered before training, fixes the first two outright (fallback 24% to 2.5%, leak 3.5% to 0.0%, both against a flat random-reward control) and moves the third (+9.1pp), while the Greek reasoning habit survives an accuracy-only gradient untouched. We release five checkpoints. The instruments, the controls and the pre-registration travel to any low-resource language; Greek is the case that let us measure them.
Ksenia Merzlyakova, Sebastian Padó, Franziska Weebercs.CL
Gender bias or other social biases in large language models (LLMs) are frequently evaluated with question answering or survey benchmarks where the LLM needs to give a response in a predefined answer format. It is well known in survey science that the answer format has a substantial impact on answers, just as LLMs are sensitive to the prompt wording. However, to our knowledge it has not been studied yet how changes in answer format impact the measurement of gender bias in LLMs and their alignment with human response distributions. We evaluate three instruction-tuned models on the BBQ benchmark and OpinionQA survey data across closed-ended, Likert-scaled and open-ended formats, comparing bias measurement and distributional alignment under otherwise identical conditions. We find that answer format does substantially alter measured outcomes, including reversals in order rankings. These differences arise because each format elicits distinct response behaviours, such as forced-choice selection, scale-based distributions and refusal in free-text generation. Our findings highlight the importance of treating answer format as a substantive component of LLM evaluation and motivate multi-format designs for more robust model assessment.
A standard claim in the literature on retrieval-augmented and memory-augmented language models is that shorter context is better when the relevant information is preserved. We test this claim by running every sample of two long-context benchmarks -- BABILong and GraphWalks (BFS) -- at four context-retention fractions (100%, 75%, 50%, 25%) under two truncation protocols. The first is the naive protocol implicitly used in much prior work: drop content from the middle of the prompt. The second is distractor-aware: identify the task-relevant content for each sample and drop only the rest. We evaluate three sizes of the Claude family (Haiku 4.5, Sonnet 4.6, Opus 4.7) and, to test cross-provider generality, GPT-5.5 from a different provider; we apply the same protocol to two further benchmarks (MRCR v2, Oolong). Under naive truncation, score collapses monotonically (paired Wilcoxon, Holm-corrected p_adj < 0.05 in all eight BABILong and GraphWalks cells). Under the distractor-aware protocol -- which preserves the signal by construction -- performance is preserved or improves: the two smaller Claude models show statistically significant gains on BABILong, while the larger models (Opus 4.7 and GPT-5.5) sit at their full-context ceiling. The naive collapse and its distractor-aware recovery replicate on GPT-5.5, ruling out a single-provider artifact. The mechanism is direct: under the naive protocol the answer-bearing content survives in fewer than 1% of samples at 25% retention; under the distractor-aware protocol it is preserved by construction. The naive protocol is therefore not a measurement of context-window effects; it is a measurement of how often middle-removal happens to spare the answer. We conclude that future studies of context-length effects must specify how they distinguish signal from distractor, or they are at best ambiguous between two opposite hypotheses.
Benchmark gains are often treated as evidence of greater LLM capability. Yet the same gain can reflect different changes in model behavior. A model may reach new answers, or produce answers that were already within reach. Aggregate scores do not distinguish these changes question by question. We establish a question-level audit under fixed budgets, temperatures, and answer formats. A question is realized when the default deployment procedure produces the correct answer. A question is reachable when a specified probe finds that answer within a fixed budget. We first test whether inference-time layer routing can expand reachability. Under a matched budget, random routes match or exceed structured search in all 43 model and task settings. Answer-blind procedures retain almost none of this gain, which instead requires access to the correct answer. We then ask why reachable answers sometimes fail to appear. Across six cases spanning 0.5B to 31B, silencing one identified MLP block repairs 68 to 92 percent of a predefined failure set. We next test whether training closes the gap by expanding reachability. In five of six matched evaluations, deployed performance rises while the reachable ceiling remains flat or falls. For DAPO, the deployed score rises by 14.7 points while the reachable ceiling falls by 13.3 points. Across the settings we audit, realization and reachability therefore do not always change together. Claims of capability expansion should report both realized performance and reachability under matched evaluation conditions. Code is available at https://github.com/LiZaiyuan0619/reachability-not-realization
Dengzhe Hou, Lingyu Jiang, Fangzhou Lin +1cs.CL cs.AI
LLM cognitive scores are increasingly summarized as per-ability profiles whose dimensions should converge across tasks, respond selectively to matched interventions, and generalize beyond the models used to define them. We introduce CogArena, a procedurally generated 13-paradigm benchmark built around a multimethod framework for determining when cognitive-task scores warrant dimensional labels across five theory-motivated groupings. Across 55 open-weight models, nearly all paradigm correlations are positive and a common axis explains about half the variance. The within-grouping advantage is small, scoring-sensitive, and uncertain across model families. In a separately frozen, fully crossed study across 12 models from six families, targeted scaffolds show a small matched-grouping advantage, but no scaffold-specific contrast survives multiplicity correction and selectivity does not improve held-out-family prediction. The frozen confirmation criterion fails. A post-hoc alternate-wording replication produces a smaller positive estimate and again fails. Together, these results support a boundary conclusion. Theory-aligned prompting produces a small in-battery diagonal tendency, but the present evidence does not establish stable five-dimensional profiles. CogArena provides a workflow joining behavioral signatures, covariance, matched interventions, and out-of-family prediction before cognitive labels are attached to model scores.
Language-model benchmarks collapse two distinct measurement questions into a single accuracy score: whether a response reached an evaluable state, and whether its answer was judged correct. We introduce a two-layer evaluation framework that separates scorer-independent execution evidence, including termination, answer exposure, parseability, and completion length, from scorer-dependent correctness. Across 2,550 outputs from five fixed Qwen and DeepSeek configurations on MATH and ARC-Challenge, matched 2,048-token limits produce sharply different execution mixtures: 49 of 450 Qwen MATH outputs terminate without a final answer, compared with 5 of 300 DeepSeek MATH outputs and none of the 750 ARC outputs. Among the same 300 DeepSeek MATH question-model pairs, no missing-final length termination is observed at 8,192 tokens. A coverage-audited targeted verification study further shows that candidate-selection and aggregation policies can substantially alter comparative accuracy estimates. These results demonstrate that accuracy conflates execution case mix with verification policy. Evaluations of test-time methods should therefore report pre-intervention execution states, verification coverage, and scorer provenance alongside accuracy.
Compressed short-text generators can fail in two different places: the codec may discard information before generation starts, or the latent generator may produce weak codes. Without separating these failure modes, researchers can spend compute improving the wrong component. We study this problem in a controlled 64-to-16 TinyStories case study built from a hierarchical VQ-VAE-2 codec and a masked discrete diffusion generator (MDLM). We use a staged validation protocol that separates codec reconstruction fidelity, latent generation quality, and auxiliary latent diagnostics under one shared external GPT-2 scorer, while reporting complementary semantic metrics for the geometry study. In the tested configuration, codec reconstruction alone raises median external perplexity from 15.17 to 27.36 (+80.4%) and p95 from 25.10 to 98.91 (+294.1%), showing that the dominant quality loss appears before latent generation begins. Under the same scorer, code-space MDLM remains materially stronger than token-space diffusion, reducing mean, median, and p95 by 32.9%, 30.9%, and 36.6%, respectively. Geometry-aware regularization improves local latent proxies but does not improve decoded-text metrics in the available runs. The contribution is methodological rather than algorithmic: the paper presents a reusable staged diagnosis for one concrete pipeline and shows that, in this setting, codec fidelity rather than latent denoising sets the practical quality ceiling.
Evaluations of language-model honesty read the model's verdicts as evidence about the model. We test the instrument instead. We built a text-adventure world where the game engine, not any model, knows whether the quest can be completed. A language model plays under a budget and must eventually declare its quest complete, unreachable, or not yet decidable; the engine scores every verdict. Decision rules were recorded before results were read, and run artifacts bind the revisions they executed; the strength of preregistration varies by series and is disclosed. With the player held fixed, instrument choices substantially changed measured behavior. On four byte-identical anchors, expanding a two-verdict grammar to three verdicts moved strong claims from 38/40 to 7/40, while the new incomplete verdict took 28/40 outcomes; across series 2, 93/158 valid games ended incomplete. One sentence disclosing the success criterion took matched-instance false verdicts from 18/59 to 0/58, through fewer decision points and cleaner decisions. Repeated runs of one fixed configuration produced non-stable verdict distributions on 3 of 4 instances: single runs report samples as dispositions. A formally preregistered narrative-register gradient was falsified; two post-hoc, hypothesis-generating patterns remain: register presence roughly doubled strong claims, and budget rendering moved verdicts more than register content (.383 meter vs .150 lantern). The narrator compressed abundant budgets toward scarcity landmarks, yet the registered mediation test returned a null. We propose a four-check integrity protocol for eval instruments.
Large language models (LLMs) increasingly issue judgments read as binary verdicts, and a growing literature reports such judgments shifting under logically irrelevant changes of wording - among them an amplified yes-no bias on moral dilemmas, absent in humans. A single framing cannot say what such a shift is: in a yes/no question the word "no" is at once logical verdict, lexical token, and last-printed option. We introduce a psychometric battery that separates these: crossed symmetrization - every logically irrelevant factor flipped in balanced pairs - across a corpus of question forms. A graded rating across logically equivalent forms recovers a coherent internal moral scale: frontier models' stance $θ$ is nearly format-invariant (cross-form incoherence 0.12-0.21 on a $\pm 1$ axis); small open-weight models fail in model-specific ways. Forcing the verdict through yes/no overlays a decomposable artifact: an order bias toward the last-printed option - opposite to classic human primacy - plus a lexical pull toward the word "no"; the artifact is substantial only in the Claude models (story-averaged -0.32 to -0.86), $\approx 0$ for GPT-5.5 and Gemini, and shrinks under extended reasoning. The word and the verdict share one token; swapping the words for arbitrary labels separates them, and the verdict-attached logical bias proves $\approx 0$ for every frontier model, while model-specific label and order attachments remain: the models are not drawn toward rejecting - the pull follows the printed surface, not the verdict it carries. A minimal model, $P = σ((θ\pm m)/s)$, summarizes any such artifact by a framing susceptibility m and a moral decisiveness s, measurably distinct from sampling temperature. The battery applies unchanged to any dilemma set and binary format: measuring what a model values requires crossing the frames of the question, not asking once.
Despite the capability of parallel decoding, diffusion large language models (dLLMs) require many denoising steps to maintain generation quality, motivating recent research on efficient decoding strategies. However, existing studies have reported inconsistent evaluation results even under seemingly identical evaluation settings, risking biased conclusions about dLLM decoding methods. To understand this evaluation concern, we conduct a rigorous evaluation of current decoding methods for dLLMs across diverse evaluation settings. Surprisingly, our analysis reveals that the ranking of decoding methods is highly sensitive to the choice of prompt templates. Single-template evaluation can lead to an illusion that decoding methods improve inference efficiency without performance degradation. Through comprehensive experiments, we find that current parallel decoding methods consistently underperform the single-token decoding baseline, failing to overcome the speed-quality trade-off. We further identify this evaluation inconsistency as the high sensitivity of parallel decoding methods to minor variations in prompt templates. Our experiments show that an effective prompt template can achieve strong evaluation results even with fewer denoising steps, markedly outperforming the marginal gain from increasing denoising steps. Beyond prompt templates, our experiments indicate that overlooked evaluation settings can also notably affect the assessment of decoding methods. Based on these findings, we propose practical guidelines for the reliable evaluation of decoding methods in dLLMs.
Style classifiers can use content cues that correlate with style labels in naturally collected data, yet we lack a systematic way to measure this reliance. We study this problem with a controlled content overlap setup built on parallel Bible translations. Specifically, we define the overlap parameter $α$ as the normalized residual of mutual information between content identity and style label, so that it measures how much content is shared across style classes: from no shared content ($α=0$) to fully shared content ($α=1$). Cross-overlap evaluation of RoBERTa-based classifiers shows that low-overlap models degrade when content cues are removed, while high-overlap models transfer more robustly. A cross-style content retrieval probe further shows that content becomes less recoverable as $α$ increases, with training dynamics showing this removal occurs gradually. Together, these results suggest that controlled overlap provides a simple diagnostic for separating style learning from content shortcuts.