The November 6 study compares stale, corrected and deleted instructions across 22 maintenance tasks and 732 runs. It tests Claude Haiku 5.5 and Sonnet 5.5 in Claude Code.
An agent is a model running a multi-step task with tools. The researchers study three real outdated instructions: a test command, a code location and a component-creation procedure.
Agents follow the obsolete test path in 78 of 80 stale-instruction runs. In 19 of Sonnet’s 40 runs, the first check after its fix executes no tests.
Piping test output through head or tail hides the failing exit status. Sonnet averages 7.8 calls to tools with the stale command versus 5.8 with the corrected command, using per-task medians.
Agents recover from 130 of 133 encounters with obsolete paths, but none of the 732 runs edits an instruction file. The requested task was fixing code, rather than maintaining documentation.
The replication package includes transcripts and analysis scripts. Correcting the instructions eliminates obsolete-path operations; deleting them leaves some agents guessing the same wrong paths.
Caveat: These are repeated runs of three cases, and the exploratory extra-call finding loses statistical significance after correcting for multiple comparisons.
(arXiv 2611.09802v1)
Retries can keep an overloaded model server stuck#
The November 5 paper reports 137 experimental runs with three models on one H100 GPU type using vLLM 0.30.0. Prompts and client traffic are synthetic.
A 107-second burst sends requests at 2.7 times measured capacity, followed by traffic at 60% of capacity. Clients allow 120 seconds per attempt and up to three retries.
With retries enabled, cancellation alone fails to restore service during the 720-second recovery window. Only about 12–38% of subsequent calls receive complete answers within their 840-second delivery limit.
In detailed Qwen2.5-7B traces, eventually aborted requests consume 79–85% of computed prompt tokens, units of information a model processes. The normal prompt counter misses roughly one-third to one-half of that total work.
Attaching retries to their original running request delivers every burst and recovery-period call, some late, when cancellation is disabled. A separate Qwen test preserves that benefit by delaying cancellation for 10 seconds.
Caveat: The delayed-cancellation tests never exercise cleanup after a client leaves permanently, and the study covers one server version and synthetic workloads.
(arXiv 2611.08684v1)
Testing what poisoned memories do in later sessions#
The November 6 benchmark, a fixed set of tasks for comparing models, contains 155 task sequences across four domains. It tests A-MEM, Graphiti, Letta, Mem0 and Memora.
Attackers alter one tool result, then ordinary tasks continue. The study separately tracks whether poison is stored, retrieved, copied into newer memories and followed in later sessions.
On 93 same-domain sequences with GPT-5.6-Luna, attacker tuning raises the share of later tasks exhibiting the target behavior from 15.30% to 46.26% for Letta. A-MEM rises from 21.02% to 44.01%.
Existing benign history changes the outcome sharply. In the separate history-prefixed, mixed-domain setting, Letta shows no later-task attack success, while Memora reaches 5.32% of later tasks.
Each integration uses its own retrieval policy and context allowance. A judge model, a model grading another model’s output, classifies poison in memory; the paper says code and data will be released later.
Caveat: These are sandboxed comparisons of complete integrations, so the percentages are neither real-world compromise rates nor a universal ranking of memory architectures.
(arXiv 2611.09721v1)
Code-review judges need better evidence as well as prompts#
The November 6 paper describes an internal Meta system that rewrites a judge’s instructions. A teacher model sees the outcome and discards examples it cannot explain from the judge’s inputs.
On an 85-task test set with execution-based labels, the rubric judge’s chance-corrected agreement rises from 0.36 to 0.60. The judge also sees the reference patch; roughly 5% of tasks are excluded per replay for errors or timeouts.
Without reference patches, the teacher abstains on 12 of 13 failed tasks in the validation set. Their defects are hidden in tests that the judge cannot see.
On 17 independent examples from the current production workflow, agreement rises from 0.16 to 0.51. On 16 older examples, it falls from 0.50 to −0.02, showing why the current workflow matters.
Filtering deliberately corrupted review outcomes reduces label noise from 24.9% to 20.4%. It still retains 19 of 22 failing changes incorrectly labeled successful.
Caveat: The production test is small and comes from one agent, while merged or abandoned changes remain imperfect substitutes for code quality.
(arXiv 2611.09291v1)
The actual release is 0.7.4.dev1, published November 4, rather than a stable 0.7.4. The project directs stability-sensitive users to 0.7.3.
Prefill is the model’s first pass over a prompt, before it writes anything. The helper runs early layers while the main Mac keeps the full model and generates the answer.
For a 64,000-token DeepSeek V4.1 Flash prompt, developers report 1,060 tokens per second on an M3 Ultra with 512 GB. Adding an M5 Max with 128 GB raises that to 3,123; a 64 GB helper reaches 1,466.
The setup guide requires matching versions and remote login on the helper. Thunderbolt or fast Ethernet carries the work; prompts below 2,048 tokens and image prompts stay on the main Mac.
The helper changes prompt-processing speed, while answer-generation speed stays the same. Its default lower-precision calculations can change numerical results, so output checks remain necessary.
Caveat: This is an experimental development build with developer-reported measurements on specific hardware, not an independently verified general speedup.
(Bluesky @macbreeze.bsky.social · collected discovery · primary release and setup guide)
Llama-modchip’s large-model CPU claim → UNVERIFIABLE: code and downloads exist, but the checked model card and README disagree on architecture details. The published three-way quality gate tests the smaller Nemotron model, leaving the compressed DeepSeek quality claim insufficiently supported here.
The claim that Opus has been weakened → HYPE: the collected discussion provides no controlled evaluation checked in this run that separates a model change from settings or workload differences.
KetchUp’s browser modeling tool → OFFSTACK: a general 3D modeling application is outside this edition’s AI-engineering focus.