One question drives my work: does an RL-trained agent behave consistently enough for its guarantees to mean anything?
Methods like conformal prediction and calibration promise statistical guarantees on agent behaviour. But they quietly assume the agent behaves consistently across training runs. If two runs of the same system — identical except for the random seed — produce qualitatively different agents, then a guarantee computed on one run says little about the other.
I trained the same multimodal agent (LLaVA-1.5-7B, LoRA, REINFORCE) on the same ScienceQA data with three random seeds, under two reward designs. With sparse binary reward, one seed produced a healthy agent and another collapsed into degenerate token repetition — the cross-seed confidence interval extends outside the possible range [0, 1]. With a continuous semantic reward (cosine similarity between sentence embeddings of the output and the reference), all three seeds converged: 6.8× less variance under otherwise identical conditions.
Sparse binary rewardσ = 0.129
Continuous semantic reward (mine)σ = 0.019 · 6.8× tighter
Same model, data, and budget — seeds {42, 123, 7}. Binary reward gives zero gradient on ~85% of training steps, so the outcome depends on luck. Continuous feedback keeps every step informative. I call the failure mode reward density collapse.
ACCEPTED · ICML 2026 AgenticUQ Workshop · Sole author
Sparse Binary Reward Degrades Agent Uncertainty Quantification in Multi-Seed Evaluation
Scaling the evaluation to 8–10 seeds and more model–benchmark pairs, to confirm or refute the reward density collapse hypothesis. Deceptive compliance — connecting my blindspot evaluations to reward design: when does a reward teach an agent to look correct rather than be correct? And representation-level grounding — factuality signals that read from a model's internals, not only its outputs.
If this overlaps with your group's interests, I'd genuinely like to hear from you — tegan.jegede@yahoo.com.