Post-training makes large language models less human-like
Source published ·Modelwire updated
Original coverage: arXiv cs.LG ↗·How Modelwire adds context
The development
A new benchmark reveals that instruction-tuning and reinforcement learning from human feedback, the standard post-training pipeline that converts base models into usable assistants, systematically erodes behavioral alignment with human psychology. Across model families and scales, this misalignment actually widens in newer generations despite improvements to base capabilities, suggesting a fundamental tension between usefulness and human-like reasoning. The finding undermines a common assumption in behavioral science: that conditioning models on individual participant profiles can recover human-level prediction accuracy. For researchers using LLMs as cognitive proxies and for teams building human-aligned systems, this signals that current optimization targets may be steering models away from authentic human behavior patterns.
Modelwire’s AI-generated summary of coverage from arXiv cs.LG.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The sharpest finding isn't just that post-training reduces human-likeness, it's that the gap widens with each model generation, meaning the problem compounds as capabilities improve. That directional trend is what makes this structurally difficult to patch with better data curation alone.
This connects directly to the GRPO gradient starvation paper covered the same day, which identified how binary reward signals in RL training create degenerate optimization dynamics. If the reward signal itself is structurally misaligned with human behavioral distributions, fixing training stability (as that paper proposes) may actually accelerate the divergence this benchmark documents. The Bayesian fine-tuning in projected subspaces work is also relevant here: uncertainty quantification during adaptation could, in principle, flag when a model's behavioral distribution is drifting from human baselines, though neither paper makes that connection explicitly.
Watch whether the Psych-201 benchmark gets adopted by any major post-training team as an evaluation gate within the next two release cycles. If it doesn't appear in a model card or alignment report by end of 2026, that's a signal the field is treating this as a research curiosity rather than a deployment constraint.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.LG
Gradient Starvation in Binary-Reward GRPO: Why Group-Mean Centering Fails and Why the Simplest Fix Works
Researchers identified a critical failure mode in Group Relative Policy Optimization (GRPO), a core algorithm for training language models on verifiable tasks like math. When all responses in a batch are correct or all wrong, the advantage signal collapses to zero, starving the policy of learning gradients. The team proved this degeneracy rate exceeds theoretical…
MentionsPsych-201 · LLMs · RLHF
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.