Repositories Become Skills as Models Manage Their Own Context
- Repository Skills Boost Research Agents. DisCo extracted over 5,000 skills from 1,000 machine learning repositories. It raised MLE-bench scores by 134.3% and PaperBench scores by 34.4%.
- Models Start Filtering Their Own Long Contexts. Declarative Attention cut attended tokens by 52.0% and 31.1% across two models. Accuracy fell by only 1.27 and 2.75 percentage points.
- EarlyEval Stops Agent Runs Early. It reduced execution steps by 13%–26% across three benchmarks. False failure predictions and late recoveries could still distort model rankings.
- Pruning Changes Persona Bias. Debias-SparseGPT adds debiasing constraints to pruning. Deployment checks now need to measure performance, sparsity, and fairness together.
Also Notable
- CRISP Selects Sparse Prefill Patterns From Input Structure — Check whether routing overhead consumes the theoretical savings.
- PaperCompiler Turns Papers Into Repository-Level Specifications — Its structured intermediate representation aims to preserve cross-file consistency and experimental protocols.
- The First Reasoning Error May Be Enough for Process Supervision — Cliff avoids training dedicated reward models for every intermediate step.
- Competitive Programming Specialization Splits Into Four Training Stages — Test the separate contributions of problem sets, synthetic traces, SFT, and reinforcement learning.
- Lens Training Data Changes Hidden-State Token Readouts — Decoded tokens cannot be treated as direct statements of a model's internal meaning.
- Optimization Agents Maintain a World Model of Edits and Scores — Belief-Calibrated Optimization aims to reduce blind trial and error from the latest feedback alone.
- Enterprise Agents Must Meet Explicit Reliability Thresholds — A higher average benchmark score does not prove a system is ready for deployment.
- Rate–Distortion Theory Puts Visual Quantizers on One Scale — The framework compares vector, product, and scalar quantization while supporting visual tokenizer selection.
- Real User Feedback Contains Signals Models Cannot Infer Alone — The challenge is separating useful information from interaction noise.
- Representation Entanglement Can Worsen Unlearning's Collateral Damage — Controlled experiments offer direct evidence for architectures designed around deletability.
Don't miss what's next. Subscribe to AI Research Brief: