Voting across AI answers smooths out random slip-ups — but does it also erase the shared blind spots every model learned from the same data?
Can aggregation cancel out independent errors but not shared training-data biases?
This explores whether combining many AI answers (voting, averaging, using several models as checkers) cleans up random one-off mistakes but leaves alone the blind spots every answer shares because they all learned from the same data.
This explores whether combining many AI answers (majority voting, averaging, stacking several models as validators) cleans up random mistakes but can't remove the blind spots that every answer shares. The corpus mostly says yes. The more interesting part is that the line between "independent" and "shared" errors is blurrier than it looks, and some training methods quietly push errors from the first kind into the second.
Start with why aggregation works at all. One result shows that a generative model trained on many imperfect experts can beat every one of them, because training on diverse data plus low-temperature sampling acts like an implicit majority vote Can models trained on many imperfect experts outperform everyone?. The important word in that finding is *uncorrelated*. The model denoises quirks that differ from expert to expert. If every expert makes the same mistake, the vote keeps it, and it can even make that mistake look more confident. The same logic sits under methods that let a model's own consensus stand in for ground-truth labels Can a model's own consensus replace ground truth labels?. That works well on benchmarks, but agreement is only a good teacher when the samples are wrong in different ways. When they're wrong together, the consensus teaches the error.
The obvious fix is to get your errors from different sources, for example by using different model families. A note on agentic validators argues this helps less than you'd hope. Switching model family closes only two of eight channels through which errors become correlated. Shared prompts, shared retrieval sources and shared provider infrastructure still make the "independent" checkers fail together Does model diversity actually reduce validator agreement failures?. Shared training data is one channel among several. Independence has to be engineered and measured, not assumed because the logos differ.
There's a less obvious twist: training itself can wipe out the diversity that aggregation depends on. RL post-training tends to latch onto one dominant format from pretraining within the first epoch and suppress the alternatives Does RL training collapse format diversity in pretrained models?. If samples from one model all come from the same collapsed mode, voting across them is closer to asking one voice several times than polling a crowd. Group-based training can also amplify rare lucky successes into habits Do overly hard RLVR samples actually harm model capabilities?, which turns a one-off fluke into a systematic tendency.
Why does this matter beyond engineering? Shared biases are the errors that accuracy metrics hide best. One critique of "theory-free" AI shows that a highly accurate model can still encode historical prejudice and correlation-causation mistakes, and the accuracy number makes them look validated Can AI models be truly free from human bias?. Aggregation makes this worse in a specific way. Random noise cancels out while shared bias doesn't, so the final answer comes out cleaner, more confident and more consistently biased. The practical lesson: when many models agree, ask whether they had any real chance to disagree.
Sources 6 notes
Models trained on diverse experts converge on consensus behavior that outperforms individuals. Low-temperature sampling concentrates outputs on this majority-voted consensus, denoising uncorrelated biases and errors across the training set.
Unsupervised on-policy self-distillation using the model's own majority-vote consensus matched or surpassed supervised methods on five benchmarks. The key mechanism distills only on self-inconsistent rollouts, using agreement as the teaching signal rather than external labels.
While different model families address two of eight shared fault channels, correlated epistemic errors likely persist through prompts, retrieval sources, and provider infrastructure. Measuring error correlation across validators differing on one channel at a time could quantify the effect.
Controlled experiments show RL consistently amplifies one format distribution from pretraining within the first epoch while collapsing alternatives. The winning format depends on model scale, not necessarily performance, and is largely hidden when starting from proprietary pretrained models.
Training on nearly-impossible problems causes models to learn degenerate shortcuts rather than genuine reasoning, and these shortcuts contaminate pre-existing capabilities. Group-relative normalization treats rare accidental successes as high-advantage trajectories, reinforcing answer repetition and computation-skipping instead of sound reasoning patterns.
Show all 6 sources
Research shows that 'theory-free' AI models mask bigotry behind high accuracy metrics while committing fundamental statistical errors. A 95% accurate criminal justice system would wrongly convict thousands, demonstrating that model sophistication does not validate causal inference.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Mechanistically Interpreting the Role of Sample Difficulty in RLVR for LLMs
- Reinforcement Learning for Reasoning in Large Language Models with One Training Example
- Sharpening Tax in Post-Training
- Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs
- Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data
- On-Policy Self-Distillation without Any Supervision
- Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining
- The Return of Pseudosciences in Artificial Intelligence: Have Machine Learning and Deep Learning Forgotten Lessons from Statistics and History?