arXiv subjects

Yixi Li

Publications and source records attributed to Yixi Li.

7 recordsLinked to original sources

arXiv2026

TxnQA: Benchmarking Analytical Reasoning and Evidence Grounding over Transaction Event Sequences

Analyzing merchant transaction histories requires dynamic temporal filtering, entity resolution, and aggregation over irregular event streams--operations poorly captured by static tabular or time-series forecasting benchmarks. We introduce TxnQA, an analytical QA benchmark comprising 27 task families (81 templates) alongside a 5-scenario fraud-risk reasoning suite that evaluates both verdicts and supporting evidence over synthetic, privacy-mitigated transaction data. Ground-truth answers are computed programmatically, and model performance is decomposed into format-following rate (FFR), conditional content accuracy (CA), and end-to-end correctness (E2E). Evaluating prompted proprietary baselines and six LoRA-adapted open-weight models shows that adaptation reliably enforces output schemas but leaves arithmetic bottlenecks: Gemma-4-31B-IT and Qwen3-8B attain E2E scores of 0.867 and 0.816, whereas evaluated MoE checkpoints reach only 0.632-0.713. Furthermore, on the 500-instance fraud suite, verdict accuracy frequently exceeds 80%, yet joint accuracy (requiring exact evidence attribution) collapses to at most ~23%. These findings identify schema compliance, semantic calculation, and evidence attribution as three distinct bottlenecks in transaction-sequence QA.

cs.AI↗

arXiv2026

A Self-Improving Agent that Grows its Own DSL: Oracle-Free Operator Induction for Exact Question Answering

A self-improving agent has to decide which new skills to keep, and in practice it keeps whatever earns a high reward, convinces a judge model, or compresses its library, even when the skill is wrong. That is fine until the answers have to be exactly right: over a table of transactions, a total is either right or wrong. We ask a harder question: can an agent grow what it can compute with no reward, no judge, and no answer key, and still be exactly right? A single frozen model runs two nested loops over a small, exact domain-specific language (DSL). The inner loop answers new questions only when two independently written expressions agree on every record. The outer loop discovers its own missing capability and authors an operator for it, admitted only when an independent second implementation agrees and the result obeys a conservation law, all without ground truth. The inner loop roughly triples the answerable-question set, and the outer loop self-extends on disjoint starting sets, inventing operators outside the starting closure; a no-induction control answers none of the new questions. The admission rule is empirically sound: it rejects 138 of 138 buggy operator variants under an independent reference, even one written by a different model family, where the admit-if-it-runs, self-consistency, and compression signals the field relies on admit most of them. Unmodified, the same mechanism self-extends on two further domains with unrelated schemas, business-ledger and clinical-encounter question answering, growing their answerable sets more than tenfold and nearly twentyfold. The loop transfers across model families too: run with further proposers, the gate stays sound and independent families self-extend comparably, with only the induction yield tracking each proposer's DSL fluency.

cs.AI↗

arXiv2026

Replaying the Movie in Your Mind Decoding Dynamic Visual Perception and Recall from EEG

Decoding dynamic visual content from EEG has advanced rapidly for visual perception, yet it remains unclear which aspects of neural representations are preserved when people internally replay dynamic scenes. We introduce PAIR, an EEG benchmark collected from 20 participants using an interleaved, clip-wise paradigm that pairs naturalistic video viewing with immediate visual recall. This design enables a comparison between externally driven perception and internally generated recall for the same stimulus content within a fixed trial sequence. We benchmark multiple decoding models on hierarchical semantic classification at both fine-grained and coarse-grained levels, meta-feature decoding tasks covering motion, color, object number, face, and human presence, and a state-decoding task that discriminates perception from recall. Within-subject validation results show numerically similar performance for several model--task combinations, alongside higher perception accuracy in two coarse-grained comparisons after multiple-comparison correction. In contrast, perception and recall are highly separable at the state level under the present validation protocol. State accuracy remains high after alpha-band ablation, although this does not exclude ocular or other task-related contributions.

cs.CV↗

arXiv2026

ATOBench: Tracing How Autonomous Penetration-Testing Agents Verify Vulnerabilities When Target Evidence Lies

Autonomous penetration-testing agents rely on target responses. These responses guide both subsequent actions and the final report. A deceptive response can therefore redirect both the attack trajectory and the agent's verification process. However, final reports reveal little about how an agent interprets conflicting evidence, changes course, decides to stop, or turns observations into a vulnerability claim. We introduce ATOBench, an evaluation framework that makes this verification process observable. ATOBench injects registered response transformations at runtime and pairs each transformed episode with a native episode under the same environment. Each pair is aligned at the first affected response. A source-linked reconstruction then follows later actions, evidence recovery, stopping, and report support. Three frozen observation contracts cover exploit proof, resource ownership, and reusable artifacts. We evaluate five model routes over 450 episodes. The analysis shows that increased activity can mask a broken verification chain, while successful recovery depends on finding usable evidence and preserving it through reporting. ATOBench turns deceptive target observations into a reproducible probe of evidence handling in autonomous penetration testing. This process-level view extends offensive pentest agent evaluation beyond final outcomes by revealing how untrusted observations shape actions, verification, and reporting.

cs.CR↗

arXiv2026

Toward Scalable Verifiable Reward: Proxy State-Based Evaluation for Multi-turn Tool-Calling LLM Agents

Interactive large language model (LLM) agents operating via multi-turn dialogue and multi-step tool calling are increasingly used in production. Benchmarks for these agents must both reliably compare models and yield on-policy training data. Prior agentic benchmarks, such as tau-bench, tau^2-bench, and AppWorld, rely on fully deterministic backends, which are costly to build and iterate. We propose Proxy State-Based Evaluation, an LLM-driven simulation framework that preserves final state-based evaluation without a deterministic database. Specifically, a scenario specifies the user goal, user/system facts, expected final state, and expected agent behavior, and an LLM state tracker infers a structured proxy state from the full interaction trace. LLM judges then verify goal completion and detect tool/user hallucinations against scenario constraints. Empirically, our benchmark produces stable, model-differentiating rankings across model families and inference-time reasoning efforts, and its on-/off-policy rollouts provide supervision that transfers to unseen scenarios. Careful scenario specification yields near-zero simulator hallucination rates, as supported by ablation studies. The framework also supports sensitivity analyses over user personas. Human-LLM judge agreement exceeds 90%, indicating reliable automated evaluation. Overall, proxy state-based evaluation offers a practical, scalable alternative to deterministic agentic benchmarks for industrial LLM agents.

cs.AI↗

arXiv2026

FLARE: Agentic Coverage-Guided Fuzzing for LLM-Based Multi-Agent Systems

Multi-Agent LLM Systems (MAS) have been adopted to automate complex human workflows by breaking down tasks into subtasks. However, due to the non-deterministic behavior of LLM agents and the intricate interactions between agents, MAS applications frequently encounter failures, including infinite loops and failed tool invocations. Traditional software testing techniques are ineffective in detecting such failures due to the lack of LLM agent specification, the large behavioral space of MAS, and semantic-based correctness judgment. This paper presents FLARE, a novel testing framework tailored for MAS. FLARE takes the source code of MAS as input and extracts specifications and behavioral spaces from agent definitions. Based on these specifications, FLARE builds test oracles and conducts coverage-guided fuzzing to expose failures. It then analyzes execution logs to judge whether each test has passed and generates failure reports. Our evaluation on 16 diverse open-source applications demonstrates that FLARE achieves 96.9% inter-agent coverage and 91.1% intra-agent coverage, outperforming baselines by 9.5% and 1.0%. FLARE also uncovers 56 previously unknown failures unique to MAS.

cs.SE↗

arXiv2026

Ada-RS: Adaptive Rejection Sampling for Selective Thinking

Large language models (LLMs) are increasingly being deployed in cost and latency-sensitive settings. While chain-of-thought improves reasoning, it can waste tokens on simple requests. We study selective thinking for tool-using LLMs and introduce Adaptive Rejection Sampling (Ada-RS), an algorithm-agnostic sample filtering framework for learning selective and efficient reasoning. For each given context, Ada-RS scores multiple sampled completions with an adaptive length-penalized reward then applies stochastic rejection sampling to retain only high-reward candidates (or preference pairs) for downstream optimization. We demonstrate how Ada-RS plugs into both preference pair (e.g. DPO) or grouped policy optimization strategies (e.g. DAPO). Using Qwen3-8B with LoRA on a synthetic tool call-oriented e-commerce benchmark, Ada-RS improves the accuracy-efficiency frontier over standard algorithms by reducing average output tokens by up to 80% and reducing thinking rate by up to 95% while maintaining or improving tool call accuracy. These results highlight that training-signal selection is a powerful lever for efficient reasoning in latency-sensitive deployments.

cs.AI↗