Training Foundation Models with a Single Algorithm
Comments on “Training Foundation Models with a Single Algorithm”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Researchers at Carnegie Mellon and Johns Hopkins propose "unified training," which treats pretraining, midtraining, SFT and RL as one likelihood-fitting algorithm over a changing data mixture. The reinforcement learning (RL) targets build on MaxRL, and a single trainer mixes static data with RL generations. In a controlled setup (a 124M-parameter model on reasoning puzzles) the best unified configurations beat staged training on accuracy, and they retained more general capability. Results come from one small model and one task family, and staged training was compute-optimal at intermediate budgets.
Why it ranked
No. 17 of 107 with a score of 0.67. The panel rated it highest on new information (0.95) and technical depth (0.90), lowest on practical relevance (0.59). Placed No. 17, No. 13 by score alone: the fifth AI & machine learning story of the day (−0.040). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Academic paper proposing unified likelihood-based training for LLM pipelines, with controlled and real-world experiments. Substantive and novel, relevant to ML builders, though mostly research-stage with preliminary small-scale evidence.
- AI judgeJudge 3gpt-5.6-luna
- Primary research presents a mathematically unified training objective, implementation details, and controlled plus real-world experiments showing performance, retention, and continual-learning tradeoffs. It is highly relevant to foundation-model researchers and infrastructure builders, though broader claims remain preliminary.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.67total under rubric v2.3
- Significance0.69
- Reader interest0.73
- Audience breadth0.66
- Practical relevance0.59
- Substantiated0.86
- Technical depth0.90
- New information0.95
- Counts against: Promotional0.03
- Counts against: Overclaim0.27
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.73 | 0.55 | 0.78 | 0.69 | 0.23 | +0.24 | +0.165 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.60 | 0.70 | 0.90 | 0.73 | 0.30 | +0.26 | +0.191 |
| Audience breadthHow many people who build technology does it concern? | 0.67 | 0.60 | 0.72 | 0.66 | 0.12 | +0.13 | +0.086 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.55 | 0.45 | 0.76 | 0.59 | 0.31 | +0.22 | +0.129 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.96 | 0.75 | 0.88 | 0.86 | 0.21 | +0.05 | +0.043 |
| Technical depthHow technically substantive is it? | 0.95 | 0.80 | 0.94 | 0.90 | 0.15 | +0.05 | +0.045 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.90 | 0.95 | 0.99 | 0.95 | 0.09 | +0.05 | +0.047 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.05 | 0.03 | 0.02 | 0.03 | 0.03 | −0.19 | −0.006 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.39 | 0.25 | 0.18 | 0.27 | 0.21 | −0.11 | −0.030 |
| Sum of contributions | +0.670 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.670 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-03/17/"><img src="https://rubric.hrih.si/badge/2026-11-03/17.svg" alt="Ranked #17 on Rubric, 3 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 8 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenBenchmark result
In the controlled reasoning-puzzle experiments, the best unified training configurations outperform staged training in final accuracy on both Base and Extended settings.
The article gives no link or citation for it.
The article says “the best unified configurations outperform staged training in both Base and Extended”
The source says No source to compare with.
No source givenBenchmark result
Unified training achieves the best accuracy and coverage for almost all compute budgets when testing generalization to larger puzzles.
The article gives no link or citation for it.
The article says “unified training achieves the best accuracy and coverage for almost all compute budgets”
The source says No source to compare with.
No source givenBenchmark result
In the controlled setting, staged training is compute-optimal for accuracy at intermediate compute budgets, whereas unified training achieves higher accuracy at larger budgets and the smallest budget.
The article gives no link or citation for it.
The article says “Staged training is compute-optimal in terms of accuracy at intermediate compute budget, whereas unified training achieves higher accuracy at larger budgets and the smallest budget in both Base and Extended”
The source says No source to compare with.
No source givenCausal conclusion
Mixing midtraining data into pretraining keeps unified models' retention similar to the pretrain-only model, while the best staged configuration has substantially higher held-out loss and lower mean benchmark performance.
The article gives no link or citation for it.
The article says “unified training maintains similar retention as the pretrain-only model, even as the amount of midtraining data increases”
The source says No source to compare with.
No source givenGeneralization
In coding RL after math training, all unified models reach higher final accuracy on both HumanEval+ and MBPP+ than staged training despite lower initial coding accuracy, supporting unified training as a better starting point for continual learning.
The article gives no link or citation for it.
The article says “Despite lower initial coding accuracy, all unified models achieve higher final accuracy on both benchmarks and higher MBPP+ coverage than staged training”
The source says No source to compare with.
No source givenNumber
The controlled experiments train a 124M-parameter model from scratch on a 20B-token pretraining corpus with a 510M-token midtraining corpus.
The article gives no link or citation for it.
The article says “We train a 124M-parameter model from scratch using a fixed 20B-token pretraining corpus and a fixed 510M-token midtraining corpus across runs”
The source says No source to compare with.
No source givenNumber
Before RL, even the largest unified midtraining mixture reaches only 2.34% and 0.75% pass@16 on Extended and Extended generalization respectively.
The article gives no link or citation for it.
The article says “even the largest mixture reaches only 2.34% and 0.75% pass@16, respectively”
The source says No source to compare with.
No source givenBenchmark result
At constant 2% mixing, REINFORCE achieves higher Math accuracy while MaxRL achieves higher Math coverage.
The article gives no link or citation for it.
The article says “At constant 2% mixing, REINFORCE achieves higher Math accuracy, while MaxRL achieves higher Math coverage”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 2 pts, not on front page
Recorded for context only, checked 3 Nov 2026, 05:47 UTC. It played no part in the ranking.