公开论文雷达

For sixty years, machine verification has been a major cost overhead, affordable only for exceptional artifacts. Here we report that generative AI inverts this relationship: at AI speed, machine verification is not only economical but essential to productivity --- it is the incorruptible referee that lets one person safely direct autonomous machine work at scale. In five weeks, one researcher on consumer AI subscriptions directed a small fleet of AI agents from application code, through a verified compiler and executive, to a RISC-V processor taped out on a community silicon shuttle; no proof passed through human review, and no RTL was written by a human. The working discipline --- the Salt method --- rests on a proof kernel no hallucinated proof can pass: mathematical claims travel between agents as kernel-checked artifacts, and human attention is reserved for statements, designs, and rulings. Verification is stated link by link, from the Lean 4 kernel to SAT-checked equivalence at the silicon boundary. We publish the complete accounting: theorem provenance, a pre-registered token meter, floor-bounded human time, and an error ledger whose catch numbering runs to #256 --- a monotone counter over the mathematics campaign's append-only flags ledger, maintained 2026-07-07 to 2026-07-20 (one number, #79, was never assigned; later catches are recorded un-numbered) --- against zero incorrect proofs reaching the record.

Formal proofs in Lean 4 that pass the kernel's type checker can nonetheless vary widely in quality. We introduce ProofJudge, an agentic LLM-as-judge system that scores formal proof quality along five dimensions beyond correctness: library leverage, automation fit, structural clarity, statement quality, and Mathlib conventions. We evaluate ProofJudge on a novel dataset of 218 declarations drawn from distinct Mathlib PRs. The judge agent is grounded by tool access to the commit the PR is applied to, enabling it to query the library state when scoring. A judge is considered aligned with human preferences when it rates the version of the PR Mathlib accepted above the initial version that was sent back for revision. All six judge models evaluated recover the reviewers' preference well above chance, from 80.8% to 63.5%, and two open-weight judges reach roughly 70% at a tenth of the best judge's cost. We release the judge harness, evaluation dataset, and evaluation traces as open-source artifacts to support further research.

AI coding agents are increasingly integrated into software development workflows, operating on both sides of the pull-request (PR) process: AI authoring agents create or modify PRs, while AI reviewers evaluate them. This creates a closed loop in which one AI coding agent reviews a contribution attributed to another. We construct a large-scale dataset of AI-to-AI code review by linking AI-attributed PRs with AI-attributed review events from CodAGE, a public dataset of coding-agent-generated GitHub events. Our dataset contains 248,641 unique AI-attributed PRs that received at least one AI-attributed review. Of these, 45,269 received cross-product review and 208,145 received same-product review; 4,773 PRs received both. Cross-product AI-to-AI review occurred in approximately 1.6% of identified agent-authored PRs but was substantial in absolute terms, and its volume increased by more than two orders of magnitude from 2025-Q1 to 2025-Q3. Reviewer output varied across author-reviewer configurations. CodeRabbit labeled 35.0% of its comments on Claude Code-authored PRs as refactor comments, compared with 10.5% on Copilot-authored PRs, although this difference may reflect characteristics of the PRs rather than the reviewer. For three of four dual-role reviewers, mean comments per PR were 58-65% higher in the same-product group, although effect sizes were small or negligible and the difference was concentrated in the upper tail. Among pairs with complete, nonnegative timestamps, the observed median latency was 1.2 minutes for cross-product pairs and 4.7 minutes for same-product pairs; differential timestamp availability and reviewer composition limit this comparison. Overall, closed-loop AI-to-AI review is increasing but remains a minority of identified agent activity, with review output varying across authoring-agent groups and product configurations.

Background. Requirements inspection (RI) is a well-established practice for detecting potential defects in requirements artifacts early in the software lifecycle. Recent advances in large language models (LLMs) have stimulated interest in their potential to support requirements engineering (RE) tasks. However, empirical evidence on the effects of LLMs when used as collaborative assistants in human-performed RI remains scarce. Aims. We aim to investigate the impact of LLM support on human-performed RI, considering inspection effectiveness in terms of smell identification and severity classification (i.e., nocuous vs innocuous), as well as inspection duration. Method. We conducted a controlled crossover design experiment with 34 participants, who inspected textual specifications with and without LLM support, identifying and classifying requirements smells while recording inspection time. We analyzed the data using one Bayesian regression model per outcome variable, accounting for validity threats induced by the crossover design as well as covariates and mediators. Results. Results show that LLM support negatively affects smell detection accuracy but has no significant effect on smell classification or task duration. A learning effect is present across experimental periods, but reduced when RI is first performed with LLM support. Conclusions. Our findings provide empirical evidence that LLM support does not necessarily improve performance and may, instead, hinder it for novice inspectors. Moreover, the results suggest that learning RI with LLM-support from the beginning may slow down the skill acquisition process, implying threats for LLM-supported learning.

Large Language Models (LLMs) deployed through Conversational User Interfaces (CUIs) are transforming health information-seeking by offering immediate, interactive experiences compared to traditional search engines like Google. However, how trust is influenced by both the types of search agents and the interface used to disseminate the information remains underexplored. This research integrates two mixed-methods studies (lab sessions and interviews) to comprehensively explore trust perceptions in health information across different search agents and dissemination interfaces. In Study 1 (N=21), we investigated trust in health information sourced from ChatGPT and Google across three types of health-related search tasks. Results showed significantly higher trust in health information from ChatGPT, highlighting the promise of LLM-powered conversational search. Building on this, Study 2 (N=20) extended the investigation to explore how the dissemination interface influences trust in LLM-sourced health information by comparing three interfaces: text-based, speech-based, and embodied, all sourcing from the same LLM. Findings revealed significant trust variations across the dissemination interfaces. Interviews from both studies revealed key factors influencing trust in LLM-powered conversational search, including source credibility, participants' search autonomy, and prior knowledge as well as the interaction style and modality. Our findings highlight the potential of LLM-powered conversational search to transform health information-seeking, underscoring the interplay between the credible search agents and the thoughtfully designed dissemination interfaces in shaping trust. These insights are crucial for developing effective, trustworthy LLM-powered health tools to enhance the health information-seeking experience.

Persistent memory makes false information durable: once a false statement is stored, it can be retrieved into future sessions that match it. We measure the cost of this failure mode using plainly worded false assertions generated in a single pass, with no instruction, trigger, or retriever optimization. Poisoning 1.2% of a LongMemEval corpus reduces accuracy from 0.850 to 0.300. A four-stage write-time screening pipeline that reaches 0.832 recall on indirect prompt injection while flagging 1.5% of trigger-word-laden benign text rejects 0 of 360 poisoned memories. We argue this exposes a boundary of content-only screening: distinguishing a false assertion from a true one generally requires external grounding beyond the text itself. We then evaluate provenance-weighted retrieval. The shipped weight is statistically indistinguishable from no defense (p=0.80), while a stronger weight recovers utility only by excluding untrusted content. In a mixed-provenance corpus where untrusted content is mostly benign, accuracy rises from 0.3167 to 0.7000; when the answer-bearing evidence itself arrives untrusted, evidence recall falls to zero and accuracy to 0.0417. Under the measured similarity regime, the additive provenance term has no usable setting: a weight strong enough to resist query-shaped poison is also strong enough to suppress legitimate untrusted evidence. We therefore argue for bounded occupancy constraints at retrieval rather than additive provenance penalties, and release the harnesses, corpora, and aggregate run reports.

Open-web future event prediction requires agents to distill reliable signals from noisy, redundant, and incomplete evidence. Existing retrieval/memory mechanisms directly feed retrieved information to agents or rely on simple memory functions such as storing and reusing prior information for prediction, leaving them insufficient for open-web forecasting. We propose to transform raw web evidence into structured memory before prediction, enabling agents to reason over distilled, question-specific evidence rather than noisy retrieval results. This paper presents ForeDreamer, a self-evolving dual-agent framework for managing memory over open-web evidence. ForeDreamer separates factual memory, a question-specific evidence state for the current forecast, from experiential memory, persistent agent experience accumulated across forecasting episodes. It uses a main agent for search and prediction, and a memory-processing subagent to convert search results into factual memory with dedicated tools. ForeDreamer further evolves experiential memory through two tracks, improving both forecasting decisions and factual-memory construction. Experiments on Prophet Arena and FutureX demonstrate the effectiveness of ForeDreamer. Project page: https://zhongzero.github.io/ForeDreamer

DreamBench-SWE is a multi-session benchmark for software-agent memory hygiene in which later software tasks depend on non-inferable evidence from earlier sessions and are scored by executable hidden oracles. We report the original scaled v2 fold and a separately preregistered v2.1 successor audit designed after that study but frozen before successor outcome inspection. The successor run completed 360/360 work units and 720/720 S3 cells across four conditions. In the original fold, the primary DF-hybrid--B5 contrast was null (95/180 versus 89/180; clustered p=.518, Holm p=1), not evidence of equivalence, and C9/C10 retained B0-headroom limitations. In the successor, no external memory achieved 21/180 passes (rate 0.1167), deterministic verbatim event memory 82/180 (rate 0.4556), the typed-plus-raw reference probe 83/180 (rate 0.4611), and one pinned hosted Mem0 literal-storage configuration 97/180 (rate 0.5389). The registered six-slot Family A retained unavailable slots at p=1; all three available comparisons against no memory rejected after Holm correction. Both preregistered mechanism contrasts were unavailable after pre-evaluation conformance rejection. The secondary literal-storage-versus-verbatim comparison was nonconfirmatory and sensitivity-dependent, while the comparison with the reference probe did not reject. The audit therefore supports DreamBench-SWE as a discriminating executable profile benchmark and characterizes one exact hosted-memory configuration, but it does not establish an external-system mechanism, superiority among memory-bearing conditions, equivalence, or broad product generality. The original v2.0.5 findings and artifacts remain unchanged.

Large language model (LLM) agents have demonstrated the ability to solve multi-step tasks requiring planning, tool use, and external information access, yet growing execution histories increase inference cost and expose reasoning to outdated, irrelevant, or misleading information, potentially degrading reasoning quality. Existing memory approaches organize or compress execution histories but provide limited mechanisms for deciding which memories remain active. We introduce the, a hierarchical memory system that organizes execution into tasks, subtasks, and actions while assigning each memory a dynamic retention score. Event-based updates and selection-based decay revise these scores, allowing WMT to preserve useful information, fold completed trajectories, suppress low-utility content, and retain access to folded context. We evaluate WMT on GAIA-Text using Qwen3-8B, Gemma 4 E4B, and Llama-3.1-8B, with ablations and memory-poisoning experiments. Relative to linear memory, WMT improves accuracy by an average of 9.97 percentage points while reducing prompt-token usage by 32.8%. Memory-poisoning experiments show that WMT limits the persistence and propagation of unreliable information. Our results suggest that effective long-horizon agent memory depends less on storing more information than on deciding which information should remain active.