Stefan Krsteski
Publications
Synthetic Persona Pretraining: Alignment from Token Zero
As language-model-based AI is increasingly deployed in autonomous settings, aligning its goals and values with those of humans becomes critical. Today, alignment, and the assistant identity itself, are typically introduced only after pretraining, once behavioral priors are already established. This can make values a thin overlay, rather than deeply rooted, and facilitate subsequent misalignment. Pursuing a different paradigm, we introduce Synthetic Persona Pretraining (SPP), which installs the desired assistant persona from token zero in pretraining. First, we annotate pretraining documents with value-aligned first-person reflections derived from a normative value constitution. Second, we pretrain via the standard cross-entropy loss on standard pretraining documents as well as their reflections, which installs the desired persona among a multitude of other personas. Finally, we post-train on user-assistant dialogue data, which binds this desired persona to the assistant identity, a process we call persona binding. By pretraining models up to 3B parameters on 500B tokens, we show that SPP improves constitution following and jailbreak robustness, and reduces the misalignment rate in out-of-distribution moral dilemmas, while preserving capabilities. Early intervention matters: compared with alignment from token zero, introducing SPP only at the end of pretraining yields weaker constitution adherence, does not shift value priorities, and leads to less aligned choices in dilemmas. This advantage depends on persona binding and, importantly, increases with pretraining budget. Overall, our results show that shaping values early is critical for alignment and establish pretraining-time persona interventions as an effective approach to do so.
Predicting Task Difficulty Without Rollouts
Task difficulty dictates an agent's likelihood of success, and estimating it without rollouts means forecasting this directly from a task description before executing costly simulations in stateful environments. Reliable estimates would therefore allow environment designers to calibrate evaluation benchmarks and construct progressive training curricula. This becomes increasingly important as agents move into long-horizon domains, where empirical trial-and-error is a severe computational bottleneck. Prior work on early prediction is limited to static tasks or isolated coding environments, often relying on narrow features and inaccurate evaluation metrics. We study \textit{ex ante} difficulty prediction across 17 agentic benchmarks spanning coding, mathematics, machine learning, web navigation, function calling, and other domains. We show that AUC can mask poor difficulty estimates, identify token-level entropy as a useful predictive signal, and show how residuals between expected and observed difficulty can expose hidden environment flaws such as contamination and infeasibility.
Messier: A High-Resolution Corpus for Cross-Benchmark Agent Evaluation
Evaluating AI agents in interactive environments is hindered by fragmented tasks, scaffolds, verifiers, and scoring rules. Existing efforts focus on narrow settings, remain limited in scale, or require costly reruns, leaving much of the empirical record incomparable. We introduce Messier, a unified corpus of 957,253 records that span 30 benchmarks, 714 agents, 11,891 tasks, and 74,205 verifiers. Messier consolidates public benchmark scores and supplements them with five-agent runs across six underrepresented professional and scientific domains, including a recent legal benchmark. Each record is standardized by model, scaffold, environment, task, verifier, and aggregation rule, with SOC/NAICS classifications for occupational and industry analysis. Using this corpus, we show frontier progress is uneven across benchmark types, with "function calling" saturated, "programming" improving the fastest, and "enterprise workflows" remaining the most challenging. Furthermore, counterfactual rescoring shows that strict all-pass aggregation in multi-verifier tasks can obscure progress and artificially alter agent rankings. From these standardized records, we derive capability scales that align with Epoch's Evaluation Capability Index rankings at Spearman \r{ho} = 0.81 and can be specialized by domain, occupation, action space, or verifier type. Messier provides a foundational, reusable infrastructure for agent capability scaling, benchmark auditing, and fine-grained analysis of evaluation failures.