Initial triage of 48 prior Agent retrievals (2026-10-07)¶
These are neither 48 implementations nor 48 completed full-text reviews. The 48 came from the prior requested window, October 2–6. Scanning also overlapped October 1, yielding 117 unread titles across the extended window. The older report placed the windows together, making them easy to confuse; original records are intact.
Each title and abstract was read. “Full text” appears only where the method was actually read. “Full-text review pending” retains candidate status and is not a final rejection or implementation.
| Paper | Review depth | Disposition in this batch | Reason or missing evidence |
|---|---|---|---|
| 2610.03675 · FrugalEvo: Towards Cost-Aware LLM-Guided Program Evolution | Full text and code | Implemented in this batch | Program evolution with strong and weak models. Real GPU experiments and negative outcomes are recorded on the detail page. |
| 2610.03631 · NeutronGym: Physics-Graded Neutron Instrument Design for LLM Agents | Abstract | Environment candidate | Requires McStas physics simulation and a hidden scientific verifier; not integrated. |
| 2610.03620 · UniIntervene++: An Adaptive Intervention Agent for Efficient Real-World Reinforcement Learning | Abstract | Environment candidate | Real-robot intervention and options; needs physical verification or a matching simulator. |
| 2610.03604 · Mastering Atari 2600 Games with Discovered Options | Abstract | Out of scope | Atari deep-RL options, not an LLM/tool-agent algorithm for this batch. |
| 2610.03598 · When a Correct Reward Is Not Enough: Diagnosing and Guiding PPO in an Analytically Solved Broker-Trader Game | Abstract | Out of scope | PPO diagnosis for financial control, not LLM post-training in this batch. |
| 2610.03591 · HazardWeaver: Scientific Route Selection for Hazard Analysis Agents | Abstract and author links | Full-text review pending | HazardWeaver has code, but requires domain data, scientific tools, and a multi-route verifier. |
| 2610.03585 · Threat-Preserving Representation Sensitivity in Agent-Security Benchmarks | Abstract | Evaluation lead | Measures security representation sensitivity, not a new policy algorithm ready for registration. |
| 2610.03574 · HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents | Abstract | Evaluation candidate | Multilingual and multimodal browsing; needs real retrieval, media tools, and an authorized API budget. |
| 2610.03525 · Structured Composition of Verifiable Atomic Insights for Table-to-Report Generation | Abstract | Full-text review pending | ComInsight's atomic SQL evidence and composition operators warrant review. |
| 2610.03524 · From Benchmarks to Production: A Text-to-SQL System for Complex Financial Data | Abstract | Full-text review pending | Domain-specific schema/reference databases for financial SQL; production claims are not industrial search/ads/recs A/B. |
| 2610.03476 · MobiAgent: Dual-Loop Recursive Policy Self-Improvement for Long-Horizon Mobile Manipulation | Abstract | Environment candidate | MobiAgent needs flow-matching skills, robot rollouts, and RoboCasa/BEHAVIOR environments. |
| 2610.03448 · Passing the Test You Trained On: Re-evaluating Prompt-Injection Detectors for LLM Agents | Abstract | Evaluation lead | Distribution-shift recheck for prompt-injection detection, not implementation of a new agent capability. |
| 2610.03356 · ReFract: Benchmarking Perspective Awareness in Language Model Agents with Text World Models | Abstract | Evaluation candidate | ReFract role-permission benchmark; verify public tasks and the text-world environment. |
| 2610.03315 · Lightweight, Rubric-Guided Trajectory Evaluation for Production AI Agents | Abstract | Full-text review pending | LiteTrajEval compression budgets and rubric judging; verify labeled data. |
| 2610.03226 · D2K-Bench: Can LLM Agents Turn Expert Designs into Efficient GPU Kernels? | Abstract | Evaluation candidate | D2K-Bench evaluates GPU kernel design. The original uses B200, so adapt hardware and do not claim matched results. |
| 2610.03213 · Toward SLM-based agentic task-tool intent matching | Abstract | Full-text review pending | SLM task–tool intent matching; needs data and the training protocol. |
| 2610.03195 · Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It | Abstract | Analysis lead | Source preference and debiasing experiments; not yet a standalone policy. |
| 2610.03153 · EvoRiskBench: An Evolving Benchmark for Runtime Security Risks in Workspace Agents | Abstract | Awaiting public artifacts | The abstract explicitly promises benchmark/platform release only after safety and reproducibility checks. |
| 2610.03136 · Investigating the Role of Reasoning-Language Alignment in Monolingual Retrieval-Augmented Generation | Abstract | Evaluation lead | German RAG reasoning-language ablation, not a new core architecture. |
| 2610.03102 · Ask, Relax, or Act? Evaluating Actionable Indeterminacy in LLM Preference Reasoning | Abstract | Evaluation candidate | Solver-grounded action, clarification, and constraint repair require a trusted solver. |
| 2610.03099 · Beyond Single Videos: Benchmarking and Active Evidence Seeking for E-Commerce Cross-Video Reasoning | Abstract | Full-text review pending | AdSeek actively collects evidence across videos with RL-SFT-RL; a multimodal candidate. |
| 2610.03089 · Securing Computer-Use Agents Against Branch Steering Attacks | Abstract | Full-text review pending | COBRA uses trusted branches and capability restrictions; needs an actual browser and STEER-Bench. |
| 2610.03056 · MOF-VERIFY: A Failure-Aware Agentic Harness for MOF Hypothesis Verification | Abstract and author links | Environment candidate | Scientific verification of MOF structures, literature, and MLIPs; domain tools are not integrated. |
| 2610.03055 · hacktrace: behavior-supervised detection of reward hacking during code generation | Abstract | Full-text review pending | HACKTRACE monitors with internal-state supervision; needs real traces, hidden states, and reward-hacking labels. |
| 2610.03036 · WebFovea: When the Model Is Right but the Click Is Wrong -- Reliable Round Trips for Vision-Based Web Agents on Live Websites | Abstract | Environment candidate | WebFovea repairs browser execution interfaces; requires live-web replay verification. |
| 2610.03033 · When Numbers Start Talking: Numerical Signalling and Strategic Behaviour Among LLMs | Abstract | Analysis lead | Multi-agent numeric-signaling games, not a training algorithm in this batch. |
| 2610.03020 · DyadMem: A Long-Term Memory Benchmark of How Agents Work with Users | Abstract | Evaluation candidate | DyadMem separates memory capture, update, and recall; needs public data and a full pipeline. |
| 2610.03014 · Beyond Predefined Sinks: Security-Aware Dependency Analysis for LLM Agents | Abstract | Full-text review pending | AgentSecGraph static dependencies and security context need validation on actual source code. |
| 2610.03010 · Engineering Sustainable Agents: A Systematic Comparison of Agentic LLMs for Developer Workflows | Abstract | Analysis lead | Measures energy and accuracy across agent configurations, not a new algorithm. |
| 2610.02994 · Sentry: Learning to Recover from LLM Agent Failures at Test Time | Full text and repository | Implemented in this batch | Failure-conditioned retrieval and reward-blind recovery verification. Author source was not released, so implementation is independent. |
| 2610.02970 · A Guideline-Augmented Multi-Agent Framework for Schema-as-Code Biomedical Named Entity Recognition | Abstract | Full-text review pending | GAMA uses guideline memory and code schemas for biomedical NER; a generic mini-suite cannot substitute. |
| 2610.02952 · GTDD: Generative Test-Driven Development for AI Coding Agents with Adversarial Testing | Abstract | Full-text review pending | GTDD adaptive testing and independent acceptance require a generator, trusted oracle, and fresh audit samples. |
| 2610.02951 · Dynamic Expert Pruning for Multi-Agent Systems | Full text | Implementation pending | Requires actual MoE routing, predictor training, and delta-loading; dense masking is not a substitute. |
| 2610.02928 · Discriminating Fixture Coverage in Agent-Infrastructure Verification Suites | Abstract | Engineering lead | Mutation testing distinguishes inactive inputs from failures invisible to the oracle; not paper capability implementation. |
| 2610.02925 · Positive-Unlabeled Learning for Agent Safety False Alarm Auditing | Abstract | Full-text review pending | Positive-unlabeled ranking for safety false alarms requires actual safety references and alarm pools; train without reading alarm gold. |
| 2610.02885 · PsyEvo: A Personalized Counseling Agent That Self-Evolves at Test Time | Abstract and author links | Environment candidate | PsyEvo simulates counseling and updates preferences. It needs specialized evaluation, and simulation gains are not clinical outcomes. |
| 2610.02847 · Turnover-Orthogonal Credit Assignment for Open-Team Multi-Agent Reinforcement Learning | Abstract | Out of scope | Open-team MARL credit under population turnover, not a language-model method in this batch. |
| 2610.02814 · VeriPy Source-Preserving Verification and Compatibility Checking for Python Components | Abstract and author links | Engineering lead | VeriPy formal verification through Dafny and Lean; verifier toolchain not integrated. |
| 2610.02744 · EpiWorld: Grounding LLM Policy Agents in Epidemiological World Models | Abstract | Environment candidate | EpiWorld epidemiological world models and interventions need domain data and a simulator. |
| 2610.02740 · Prospective Hindsight: Self-Calibrating Reinforcement Learning via Prediction-Reality Gaps | Abstract | Full-text review pending | Prospective Hindsight reweights prediction–outcome gaps; verify gradients and training recipe. |
| 2610.02710 · Self-Supervised Scaling of Terminal Environments for Scientific Domains | Abstract | Full-text review pending | Reconstructs scientific terminal environments with a hidden verifier; needs 500 workflows and actual execution. |
| 2610.02702 · Silent Dissent: LLM Agents That Yield to the Majority Still Represent Their Original Premise | Abstract | Analysis lead | J-lens observes latent representations after conformity and is not directly a self-improvement policy. |
| 2610.02687 · Decoupling Memory from Context: Structured Memory for Token-Efficient Test-Time Continual Learning | Abstract | Full-text review pending | GraphMemory retrieves query subgraphs and updates memory; verify update and merge rules. |
| 2610.02670 · LEAP: Learning Efficient Action Proposals For LLM Agents | Abstract | Full-text review pending | LEAP distills small-model action sequences and verifies with the target; needs actual latency and success-rate controls. |
| 2610.02654 · Coherence-Driven Belief Formation and Population Dynamics of Contagion in LLM Agents | Abstract | Analysis lead | Social belief contagion experiments, not a new general-purpose agent operator. |
| 2610.02638 · Batched Speech Decisions Without Decoding: Single-Token Supervision Lets a Frozen LLM Hear Beyond the Transcript | Abstract | Full-text review pending | DuplexJev ASR hidden-state connector and single-token decisions; a System One/multimodal candidate. |
| 2610.02617 · WebUIProof: Benchmarking WebUI Code Generators with UI-Agent Execution Harness | Abstract | Evaluation candidate | WebUIProof needs actual browser execution tests and RL; real UI training is not enabled. |
| 2610.02616 · VERSE: Verified Self-Evolving Optimizer for Agent Harnesses | Abstract and author links | Full-text review pending | VERSE uses verification-driven two-level harness evolution; requires isolated code execution and SWE tasks. |