AI Agents

CudaPerf adds structural code-aware rewards — memory coalescing, occupancy, arithmetic intensity, synchronization patterns — on top of the usual correctness-and-speedup RLVR signals for CUDA kernel generation. It runs in two stages: offline contrastive ranking of strong vs weak candidates, then online RL jointly optimizing correctness, performance, and structural efficiency, with iterative refinement from execution feedback. They release 2.9k C-to-CUDA and 1k PyTorch-to-CUDA programs with multiple implementations. The reward design is the interesting part; the truncated abstract omits the actual speedup numbers.

This benchmark measures adaptive multi-round prompt-injection attacks against memoryless defenders across 21 scenarios: the attacker observes prior defender responses and pivots, while each defender turn is scored fresh. Restricting to the first turn yields 0-1% attack success; allowing 15 adaptive rounds pushes it to 5.4-14%. Pooling three frontier attacker LLMs finds 1.4-2.2x more unique successes than the best single attacker, with attacks barely resembling existing corpora (cosine 0.02-0.14). Opus 4.6 and GPT-5.4 tie in aggregate but fail on different scenarios.

AttriMem tackles the credit-assignment problem in learned memory policies: outcome- or module-level rewards tell you the task succeeded but not which stored memory contents actually helped. It augments the global outcome reward with local rewards derived from token-level contributions to the final answer, giving fine-grained process feedback where no ground-truth intermediate targets exist. On long-horizon dialogue QA it beats retrieval, heuristic, and RL baselines and stabilizes RL optimization. The token-attribution mechanism is the contribution; the abstract reports no cost or latency figures.

A sharp result for anyone building bug-reproduction into repair pipelines: the standard fail-to-pass criterion for generated reproduction tests is insufficient. Some F→P tests are lax — they reproduce the symptom but still admit plausible-but-wrong patches, and co-generation couples test and fix errors so the in-loop F→P check passes when both are wrong. CoHarden generates the test first, then iteratively hardens test and fix against surviving mutation patches until the test stops admitting lax regressions. Only rigorous tests consistently improve repair.

This reframes automated remediation around the decision of whether to intervene, since an incorrect repair often costs more than doing nothing. It models safe remediation as a Constrained MDP maximizing repair success under a bounded false-remediation rate, and decomposes per-action risk into blast radius, reversibility, and epistemic uncertainty for an interpretable operator interface. A context-adaptive human-in-the-loop gate escalates based on on-call load and business criticality. On Train Ticket with Chaos Mesh injection: 39% lower FRR, +2.5 points repair success, 17% less escalation.

The new rules of context engineering for Claude 5 generation models

Ruff 0.16.0 flips the default rule set from 59 to 413 enabled rules, out of 968 total. That means any project with an unpinned ruff dev dependency will see CI light up with hundreds of new findings, some catching real syntax and runtime errors previously off by default. Willison ran uvx ruff@latest check . against Datasette, sqlite-utils, and LLM and found hundreds of minor issues. Relevant if you let agents auto-upgrade tooling.