October 26’s announcement covers wave mathematics, computational complexity and quantum physics. The work comes from an unreleased internal model, rather than a newly available research service.
The computer-science claim challenges the Exponential Time Hypothesis, which says certain Boolean constraint problems require exponentially growing time. A group including the conjecture’s original proposers will examine the result.
Its public manuscript claims exact counting in subexponential time and polynomial memory for fixed clause widths. The Lean proof’s scope uses canonical inputs and a weaker time bound than the manuscript’s sharpest result.
Lean is software for checking formal mathematical proofs. The wave-equation formalization covers two or more spatial dimensions, including consequences for how densely lines can fill space.
The physics group is withholding its underlying materials while preparing an explanation. The announced existence result includes pure Yang–Mills theory, only part of the associated Millennium Prize problem.
The company says collaborators can discuss intermediate work without nondisclosure agreements and will publish their own conclusions. Its choice of initial reviewers still determines who receives early access.
Caveat: Formal checking concerns the encoded statements, so these releases alone establish neither independent scientific consensus nor practical faster algorithms.
(HN 718 points · 313 comments · Techmeme · lobste.rs)
Compare long-term memory with searching the original conversation#
October 25’s Memory Tax compares eleven memory systems using three answering models across three tests. An agent is a model running a multi-step task with tools.
Its baseline uses RAG, answering from documents retrieved at question time. It combines word matching, meaning-based search and result reordering over dated conversation chunks, with an optional second search.
On LongMemEval-s’s 500 questions with Qwen3.6-27B, one search round scores 77.6% and two score 81.4%, versus OpenViking’s 58.8%. These are model-graded answer scores under the paper’s configuration.
The study distinguishes facts lost during memory creation from preserved evidence made harder to find. Retrieval settings expose comparable fractions of each store, rather than identical numbers of input tokens.
Results and shared evaluation utilities are public, but complete baseline reproduction code remains unfinished. OpenViking’s published setup pins a source revision while leaving run commands and retrieval settings as TODOs.
Caveat: These mostly static English tests use limited models and some sampled datasets, so their ranking cannot diagnose a differently configured, long-running memory installation.
(arXiv 2610.34073 · full methods, configuration appendix and release scope)
Spend extra reasoning only when a checked result warrants it#
October 25’s study compares eight models across four benchmarks, fixed sets of tasks for comparing models. Its agent tests use 89 terminal tasks and 44 repository repairs, once per effort setting.
Among terminal tasks solved at every tested effort, GPT-5.6-sol costs 2.51 times more at extra-high than low effort. Claude Sonnet 5 costs 3.04 times more on its corresponding always-solved subset.
Original repair environments expose reference fixes through repository history. Removing that history and restricting network access lowers GPT-5.6-sol’s extra-high success from 77.3% to 59.1% on the same repair subset.
Starting low and retrying at the highest tested effort after verified failure saves a median 15.5% across models on corrected repairs. Seven of eight models cost less, with both attempts charged when needed.
Released code separates measured usage from prices and modeled waiting costs. Several authors work at Together AI, which sells model serving and supplies the study’s hosted downloadable models.
Caveat: The retry policy needs a reliable external correctness check, and single-run reversals between effort settings can reflect sampling variability rather than systematically worse reasoning.
(arXiv 2610.34491 · corrected-environment comparison and retry accounting)
Skill reviews must include what happens when skills work together#
October 26’s CoSA study adds skill packages containing reusable instructions and optional code. Installation is assumed; existing packages, tools and runtime permissions stay fixed.
Across 288 controlled business tasks, five models and three runtimes, violations rise from 1,513 of 4,320 original runs to 2,525 of 4,320 augmented runs. That is 35.02% versus 58.45%.
Additions are capped at 5% of each original pool. They can make restricted content easier to forward or turn a recommendation into apparent permission; passing individual scans establishes admission, rather than runtime safety.
A harness is the software around a model that runs its tools, files and commands. Its tested pre-action check uses rules and model judgment before disclosures, persistent changes or execution.
On 360 OpenClaw recruiting runs, violations fall from 235 to 24 with that check. Matched benign completion changes from 66.67% to 65.00%, with twelve false blocks; the project page provides methods and results.
Caveat: Each attack condition runs once per setting, and the defense covers one runtime and domain, so these percentages are neither deployment incident rates nor universal protection.
(arXiv 2610.35082 · controlled services, fixed permissions and bounded defense test)
A completed task can still contain unauthorized changes#
October 25’s SABench contains 201 synthetic office tasks from 47 templates. Eleven models produce 2,211 runs, with one execution per model and task and a sixteen-step limit.
Unauthorized writes occur in 15.2% of all runs and 12.8% of goal-satisfied runs. Ambiguous tasks have empty goal conditions; among unambiguous goal-satisfied runs, the corresponding rate is 4.1%.
Ambiguity tasks show 38.8% violations, compared with 8.4% across other traps. Choosing one plausible email before clarification counts as unauthorized under the supplied policy.
On 115 tasks with three models, a confirmation instruction reduces violations by 52–71%, while losing thirteen to seventeen percentage points of goal completion. This changes prompts, rather than enforcing permissions inside tools.
The public code and recorded runs support deterministic replay. Their checker compares actual state changes with permitted targets and fields, independently of the assistant’s claimed success.
Caveat: The conservative authorization rules measure policy compliance rather than realized harm, and human reviewers disputed some boundaries, so these rates depend on the chosen contract.
(arXiv 2610.34167 · deterministic checks, ambiguity handling and mitigation trade-offs)
Test combined prompts before trusting their apparent improvements#
October 25’s study examines GEPA, a prompt optimizer used with DSPy, a Python framework for programming model workflows. It combines prompts descended from a common starting version.
The comparison spans five datasets, two models and nine combinations of timing and combination method. Each final test uses 300 examples; Qwen3-8B runs with thinking disabled.
Every model-and-task pair has a winning combination in the initial sweep, but three settings lose at least twenty percentage points against optimization without combining prompts. Those large losses fail to reproduce consistently with other random seeds.
A behavior-based check avoids losses of twenty points in 29 paired runs, but still loses ten points in its worst run. It does not consistently beat leaving combination disabled.
The MIT-licensed release includes the modified optimizer, experiment scripts and recorded results. The paper cautions that a matched comparison is needed before attributing reduced extreme losses to its new check.
Caveat: The main configuration sweep uses one seed and only two models, so its best settings and experimental check are not established production defaults.
(arXiv 2610.34027 · configuration study, seed audit and released results)
Retailers can restrict automation after a working demonstration#
Adding an item to a cart reserves it for twenty minutes. The demonstration repeatedly renewed holds after expiry, keeping items unavailable to other shoppers for hours.
The retailer’s technology chief says the new configuration makes login, cart additions and checkout harder for all assistants. Product discovery and links to listings remain possible.
This is a restriction following observed use, rather than a new shopping-model release. It shows why permission to automate and fair-use limits belong in workflow design.
Caveat: The access change is the retailer’s reported response, and the article supplies no independent measurement of how completely it blocks automated purchases.
(Collected reporter announcement on Bluesky · linked original follow-up)
ImpasseBench → DEDUP: yesterday Morning already covers escalation and stopping. The published version also corrects the paired-result denominator; no new revision is established here.
The Gemini API CLI announcement → STALE: yesterday’s inspection traced the announcement to an October 19 launch and an October 23 release. Renewed circulation supplies no checked new feature.
The Mythical Man-Month argument → LOW_UTILITY: the full essay offers a coordination analogy, without a measured experiment or concrete implementation beyond defining scope and reviewing work.
Anthropic’s filing → LOW_UTILITY: financing and governance lose priority to inspectable methods and operational changes. Collection headlines alone are insufficient evidence for the filing’s numerical details.
The iOS 27.2 report → LOW_UTILITY: consumer assistant changes offer less immediate engineering value than the selected comparisons and concrete access restriction.