Papers

A small learned controller compares proposed actions with user authority and an alternative, with important differences between classification and reversible execution tests.

VerifierBench tests whether agents can write executable task checkers. An agent is a model running a multi-step task with tools. A checker must recognize different valid solutions, including ones its author never chose. The authors report a …

Store the order and prerequisites of successful work, then retrieve compact instructions instead of loading every skill.

A two-model GEPA and SkillOpt study separates answer correctness, formatting, server failures, waiting time and token use.

SkillClash isolates failures after both skills are loaded. A correct combined example repairs much of the lost rule compliance.

SWE-Bench Pro V2 separates bad instructions from evaluation exploits. Its strongest systems already approach the repaired test set's ceiling.

Arcade compares finding tools on demand with listing 100 tools in every request. Savings depend on search frequency, prompt caching and how the client handles large results.

Munin combines local paper search, full-text reading and citation checks with a rerunnable evaluation suite. Its strongest result permits web access and depends on a private reference corpus.

Validation and a constrained retry recover tool-request errors in controlled Qwen tests. The strongest results depend on resolving a usable tool name, with weaker gains outside that model family.

An external controller learns short repository instructions, hides unused tools and redirects stalled work. Reported savings include its own model requests and prompt-cache discounts.

A memory test follows facts as they change, exposing errors that answer scores can hide or misdiagnose.

An eight-GPU serving study fixes a cache bottleneck, while retaining the faster median of its recomputation baseline.

A larger tool menu looked like worse tool use until researchers raised the model's response allowance.

A second model runs checks after completion, then sends failures back while preserving the working agent's state.

Agent Step-Up tests whether approval screens show the real action and whether that action can change after approval.

A shared interaction loop and executable task checks matter more than the nine-billion-parameter model's headline scores.

CODA tests jointly trained models and skill libraries after their instructions change. Even paraphrases can break previously successful tasks.

Slow reasoning improves a fast game-playing agent when it happens between attempts, with validation deciding which instructions survive.

A single-GPU study preserves returning sessions while loading the next expert, but its extreme capacity claims need clarification.

HybridAgent mixes internal vector-based computation with text for tools and inspection. Matched-model tests show stronger search results, but the cost savings cover estimated generation compute.

Keep the work products, their evidence and unfinished requirements, so a long-running agent can stop rereading every step.

A controlled study separates finding a skill, opening its instructions and completing the task. Choosing skills before execution helps, with costs.

A model can reject the right tool argument when its source is missing, or when using it requires an inference.

Copies of one old claim can crowd out a correction. A memory study tests source tracking, and shows when filtering makes answers worse.

A received revocation message can leave tool gateways accepting old permissions; a bounded monitor checks enforcement and actual writes.

An analysis of 600 coding-agent runs finds that repeated steps explain much more cost variation than tool-output size, under simulated cached pricing.

Encrypted networks expose a limit of capability testing: failing to uncover an ability cannot universally prove it absent from adversarially supplied weights.

ByteDance's production memory system for LLM apps: events as a log, entities as materialized views, 83% less storage than raw logs, and — the part that matters here — OpenViking is its open-sourced subset.

An NVIDIA researcher and a cartography professor formalize the context window as terrain to be mapped, then score four agent systems against seven operators and find the industry at roughly 60% coverage.

A practitioner study separates reported agent reliability from measured repository habits and the work of maintaining specs.

ContractSim separates reaching a deal from honoring it. Private earnings can reward missed payments, and changing the instructions changes the result.

A derivation-led route from PCA to diffusion, exact likelihoods and energy-based models, with a useful correction in the revised edition.

Models repeatedly recommend Japan’s Communist Party to left-leaning voters. Source access offers a possible explanation, but the pattern survives without web search.

A research sampler broadens AI-generated ideas by scoring plausibility separately from familiarity. The evidence shows a trade-off, not guaranteed discoveries.

A geometric alternative to attention trails small Transformers, and its written method leaves a serious question about access to future tokens.