Machine Learning
Research on Machine Learning: explanations, evidence, limitations and connections between papers.
79 results
A flow-wide method improves low-temperature generation
Bayes-Consistent Tempering aligns training and sampling for flow language models, improving low-temperature results in tested math and Sudoku tasks.
A memory branch improves GraphCast’s longer-range forecasts
A memory correction branch lowered GraphCast’s weighted forecast loss across ten-day forecasts from 122 starts in 2023.
Match parallel generation to the work that slows it down
A phone agent can wait at every screen action, while a text generator can spend time proposing, checking, or reconstructing its output. Parallel work helps when it reaches the costly part of that particular workflow. Across these studies, moving or overlapping work produces reported throughput or latency gains in specific workloads and serving setups. JetSpec++ also reports a serving gain, while its predictions of future context remain estimates until verification. The results support judging each method by its local outcome, without forming a shared speed ranking.
Studies test final self-improving systems more than the loop behind them
A study of papers finds repeated tests of full improvement loops less common than evaluations of final systems.
A new method tracks how training targets change predictions
The target-response Jacobian closely tracks refitted prediction shifts after training-target changes on three tested tabular datasets.
A score can rank when CT systems fail without warning
A prospective score ranked silent failure in two learned CT systems, with evidence limited to simulated access tests.
Racer checks whether GPU kernel rewrites preserve their function
Racer symbolically checks GPU kernel rewrites and exposes bugs missed by sampled tests, within the coverage of its production-kernel study.
Shared anchors let language-model monitors transfer across models
Shared-anchor probes reuse behavior monitors across language models, but transfer varies by task and target.
Two-class pretraining still supports multiclass table predictions
A TabICLv2 study finds two-class pretraining retains multiclass prediction, with results bounded by training recipe differences.
Same-size students struggle to match model harnesses
Trace-trained students lag behind harness-assisted models on tested puzzles, while greater capacity or simpler tasks can narrow the gap.
Synthetic training can fail where generated data leave gaps
A theory of synthetic training shows when coverage supports real-data error guarantees and how uncovered regions can defeat them.
A speaker’s ordinary voice can trigger a speech-model backdoor
A two-phase method ties a target voice to chosen outputs, succeeding selectively in two tested speech models.
HarnessPro improves agent success at lower cost
HarnessPro evolves AI agent software with execution feedback and generated tests; one terminal-task comparison improved success at lower cost.
Small adapters improve brain-response predictions across tested models
BrainLoRA trains small adapters to improve brain-response predictions across tested models, with transfer evidence limited to narrative films.
Text-to-image models can learn without text-only pretraining
Halcyon learns from paired captions and images; tests found no benefit from text-only pretraining in the specific settings studied.
AI weather forecasts retain skill beyond local heat records
A 2022 comparison finds AI weather systems retain heat forecast skill beyond local records, with less consistent results for wind.
Jailbreak rankings shift with evaluators and attempt settings
A study finds jailbreak attack rankings shift with evaluator rules and attempt protocols across tests of three small open-weight models.
A shared HTTP guard makes Java model calls more consistent
A shared Java HTTP guard made three client paths behave alike in scripted provider-failure tests, with a slow-stream timeout trade-off.
Compact configurations improve hardware generation in one SECDED task
A compact configuration improved formal-equivalence results over code-like outputs in one bounded hardware-generation task.
Parallel language models coordinate several token predictions at once
A 1.3B Parallel Language Model beat a same-size autoregressive model on one math test's accuracy and question rate.
Two fully trained loops can match large models with fewer parameters
A scaling law predicts validation loss for looped transformers and favors fully trained loops when both loop counts can vary.
Match routing decisions to the evidence available at each step
An agent may need to choose a model for a home-control step, decide how to ask owners of separate tables for an answer, or send a question to a larger model. Each choice can consume additional computation, yet each exposes different evidence about where that work may help. A useful first answer is to match the routing decision to the variation the system can observe: a known step, a dependency between owners, or a question-level confidence signal. These studies test those choices in different settings, so their outcomes guide conditional decisions rather than a shared performance ranking.
Match judging evidence to the outcome and its cost
A panel of judges can give more consistent scores after its instructions are clarified, yet show no detectable improvement against expert ratings. A response can also look relevant without answering the request. These cases make “better judging” an incomplete goal until the outcome is named. The studies support choosing evidence for the claim at hand: clarify criteria to study consensus, inspect an answer’s substance to assess request matching, and observe an application in use when its behavior matters. Each choice has a tested scope and a cost; a budget planner depends on how well its pilot represents deployment.
Blitz cuts protein structure sampling to eight steps
Blitz distills AlphaFold 3 and Boltz-2 sampling into eight steps while retaining checked structural accuracy in reported tests.