Papers

A review finds faster single-request generation with four-bit weights, little gain in one large-batch test, and sparse matched accuracy evidence.

Two grading models can agree on the same mistake. Cross-provider agreement filtered one unreliable grader while discarding most comparisons.

Rewriting earlier reasoning often shortens later reasoning, but the compressor can erase time savings and task quality remains uncertain.

Across 120 historical distributed-system bugs, coding assistants find relevant files more reliably than the precise code that caused the failure.

Separate working copies keep code intact, but task-specific history and progress records help coding assistants resume interrupted work.

A web measurement study finds that successful status codes can hide empty shells, bot checks and substitute pages, distorting citation checks.

Genliang Zhu and colleagues study what a tool-using agent should do after a timeout leaves a remote action’s outcome uncertain. A retry can duplicate an effect that already happened or is still in flight. Their controller considers …

Anirudh Phukan and colleagues test whether search-equipped assistants discover recent changes without being told what to check. They pair direct questions with recommendation requests that depend on the same facts. Eight models were tested …

A paper tests how different descriptions of the same scene help text and image models learn similar representations.

Loading pages can invalidate a correct browser plan; pixel checks and outcome checks solve different parts of the problem.

Tests of self-improving instructions can miss redundant protections disappearing, or conflicting instructions added around unchanged safety rules.

Written priorities redirect coding assistants more reliably than another conflicting code example in controlled repository experiments.

GAAC reconstructs compressed conversation state inside attention, trading a little computation for less memory traffic and room for another request.

Poisoned training text can leave a hidden command trigger after clean safety training; the headline rate counts 32 attempts.

A Python library gates actions by conversation state; its airline tests also expose how outcome scores miss skipped consent.

FeatureArc tests coding agents, models running multi-step software tasks with tools, as they extend one feature on their own earlier code. Each new request inherits any unresolved mistakes. The authors evaluate ten agents on 73 reviewed …

Explicit subtask boundaries improve author-run coding tests, but the cost savings vary by task and model.

A large text study finds reusable instructions that loosen human oversight, including in otherwise ordinary automation workflows.

GEPA improves prompts for model programs by proposing changes and testing their behavior. It works with DSPy, a Python framework for programming model workflows. This extension proposes several changes together so evaluations can run …

A coding agent helped researchers revive two historical systems that made programming knowledge explicit.

A controlled synthetic-memory study separates finding an update from selecting it, with limits from familiar facts and scoring.

A diagnostic agent lost a third of its tool results before the model read them, even though its largest prompt used only 5.75% of the available window. The surrounding software sliced long JSON, a structured data format, at 6,000 …

PROBE tests whether coding agents find defects that maintainers later repair. Its reference set contains 2,457 bugs fixed over eighteen months in twenty repositories, all already present in the starting snapshots. The authors split audits …

CoSA studies reusable agent skills: packages of instructions, optionally with code. Passing an individual package check can still leave unsafe interactions when one skill reuses another’s output. A recommendation can become apparent …

CacheLint checks application code for prompt-caching mistakes. Prompt caching reuses computation for an identical beginning of a request; an agent can lose those savings while continuing to answer correctly. The authors inspect 120 merged …

Kaiser Sun and colleagues compare reasoning-effort settings across eight models and four task sets. Higher effort often improves aggregate accuracy, while spending more on tasks that cheaper runs already solve. The study measures whole …

SABench tests whether an assistant finishes the requested work and whether every write stays within permission. Replying correctly to an email can coexist with an unrequested change to a linked ticket. The study uses 201 synthetic office …

Memory Tax compares eleven memory systems with search over the original conversation. It asks whether extracting and organizing memories earns its extra cost. The authors separate new answers that memory enables from answers lost relative …

GEPA evolves prompts for model programs, including workflows built with DSPy. This study tests its merge operation, which combines candidates descended from a common starting instruction. The comparison spans five datasets, two models and …

MCP-Injector inserts errors, delays and altered results between an assistant and tools connected through MCP, the standard for connecting agents to tools. It tests reliability when connected programs behave badly. The authors cross …

A refund request can commit even when its acknowledgement disappears. If permission changes before a workflow restarts, a retry needs two separate checks: current authority and evidence about the earlier effect. Arya.ai studies 336 …

ImpasseBench tests whether coding assistants stop when their own tool results show that outside help is needed. It separates persistent barriers from temporary errors and blocked testing. The authors run eight model configurations on 102 …

Upwork’s SWE-Clar asks coding assistants to clarify incomplete requests before building software. It freezes that conversation, then starts two isolated continuations: one answers requirement questions and the other writes code. Assessment …

A coding assistant can preserve its task score while forgetting restrictions on how it should work. This study compares five conversation-shortening strategies inside OpenCode, the software running the model’s tools and commands. The …

MacPaw researchers test four families of model-serving optimizations on Apple Silicon. Their measurements show why a server-GPU speedup needs checking against the model, software and request pattern you actually use. Pruning the KV cache, a …

AfterDeleteBench separates deleting a record, preventing disclosure and blocking a later write. Its controlled tests also expose failures hidden by answer parsers.

DoorDash describes a narrow coding task with explicit safety checks: remove obsolete feature flags, the settings used for gradual releases and experiments. The system queries deployed values before an engineer approves each proposed …

Yuning Bian and colleagues compare document splitting methods across 120 adjacent Wikipedia revision pairs. A small edit can shift boundaries and force a retrieval system to embed unchanged text again. On the 116 pairs all methods process …

Abhishek Barali and Aayush Paudel test what happens when an extraction schema requires more answers than a document contains. Models often fill the gap with genuine text of the wrong kind. Across eight downloadable models and 558 documents, …

An audit finds hidden interface requirements in SWE-Milestone. They distort some results, while removing affected tests largely preserves the overall ranking.

Jens Lehmann and Sahar Vahdati let models build and twice revise their own harnesses: the software that runs their tools, files and commands. Improvements on familiar development games often overstate gains elsewhere. The study follows 27 …

MemPABench asks whether an assistant remembers how a person wants help, including when to confirm, ask questions or initiate contact. It measures the preference selected and the behavior actually delivered. Four simulated users have …

Text-only editing damaged some correct transcripts. Adding audio reduced that damage for Gemini, but also reduced repairs to its own drafts.

Yikai Zhao and colleagues test whether document conversion preserves the requirements an author marked for attention. Words can survive extraction while color and highlighting disappear. Their task asks models to recover specially marked …

Zedong Peng compares extracted memories with coding agents that search raw conversation history. The study includes store-construction costs, which can outweigh cheaper answers later. It tests seven methods with three models on one LoCoMo …

Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald examine why AI prose can feel difficult despite being grammatically valid. Their study finds features associated with reading difficulty across model-generated conversations, coding …

A study of ChatGPT, Claude and Gemini shows how saved malicious instructions can affect later chats. Memory injection means storing an instruction that returns as context in future sessions. The authors tested ChatGPT Instant 5.3, Claude …

Single-block profiling finds that attention software leaves newer workstation GPUs underused, while proposed fixes remain a systems research agenda.

Anti-Persistence saves execution records and rebuilds advice for each task, reducing one route by which an agent can keep repeating a bad lesson.

AI can turn shared knowledge into private wealth while overwhelming the reviewers and maintainers who keep it alive.