31 reviewed and deferred leads: not accepted for implementation¶
The old script promoted reviewed_deferred records from phrases such as 'pending implementation,' 'next batch must,' or 'retain for next batch.'
All leads and their reasons remain below. They are neither complete nor part of the 504 structurally accepted records.
| Paper | Rationale type | Original rationale |
|---|---|---|
| 2610.13166 Scaling Agentic Data for Long-Horizon Terminal Intelligence | integrity_or_acceptance_audit_preconditions | Reviewed TerminalHorizon methods and experiments: independent execution finds precompleted work in synthetic environments before reconstruction and expansion of same-goal tasks. 1.5k environments, Qwen3.5-27B SFT, 12 benchmarks. Production settings are not production A/B. Isolate gold environment verifiers from agent-visible packages. |
| 2610.13144 Competence-Aware Orchestration for LLM-Based Multi-Agent Systems | implementation_or_execution_preconditions | Reviewed competence orchestration: fixed-probe capability fingerprints, query-conditioned prediction, learned role activation and communication graph, six reasoning/code benchmarks and unseen backbones. Train an actual competence predictor and isolate probes/test; no gold capability inference. |
| 2610.12957 JoyInside-VLM-Agent: Product-Conditioned Multimodal Policy Learning for Embodied Agents | implementation_or_execution_preconditions | Reviewed JoyInside: product-conditioned multimodal trajectories, SFT, multi-teacher OPD and DPER token-level privileged-evidence routing. The serving student does not read teacher references. Private 120k mixed records, skills, and two robot cases limit public reproduction; retain actual multistage training. |
| 2610.12794 Meta-Reflective Evolution Effectively Improves Production AI Systems | further_evidence_verification | Reviewed CoCoEvolve: archive parents, retrieve history, probe live targets, self-test mutations, then score externally. Train examples visible, validation score-only, test isolated. Production AI targets and public holdout do not imply traffic A/B. Verify live authority and actual evaluator, not registry labels. |
| 2610.12655 CLEAT: Learning Value-Guided Clarification for Interactive Agents | implementation_or_execution_preconditions | Reviewed CLEAT abstract, intent alignment, and experiments: paired ask/proceed rollouts from the same history estimate residual query value; co-train controller and policy and refresh labels with new outcomes. Qwen3-4B, four benchmarks, ten settings, three seeds. Actual predictor and policy required; no hidden test goals in input. |
| 2610.12611 SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments | implementation_or_execution_preconditions | Reviewed SoftwareGym: generate executable GUI apps and repair against frozen v1 GUI tests; 182 apps across 12 categories, 120 tasks, Qwen3.5-9B multimodal SFT and OOD OSWorld. Run real GUI interactions in a VM and isolate environment self-tests. No static screenshot or scripted-action substitute. |
| 2610.12248 EgoVoice: Proactive Spoken Assistance from Egocentric Multimodal Streams | implementation_or_execution_preconditions | EgoVoice trains proactive listen-or-speak SFT/DPO from HoloAssist source-separated/resynthesized audio, evaluating onset, content, and ending in 170 sessions. Follow-up must retain actual streaming audio/video and timing policy. |
| 2610.11862 LEVER: Adaptive Cost-Aware Proof Search Over AND/OR Graphs | implementation_or_execution_preconditions | Reviewed LEVER AND/OR search: realized partial-proof objective plus open-subgoal predictions, configurable cost/length/impurity, correctness by Lean kernel, matched model/version/price. Run actual search, score before predictions are tested, and charge failures; no gold-proof ranking substitute. |
| 2610.11578 Chronos Enables Code Agents to Reason over Software Evolution | implementation_or_execution_preconditions | Reviewed Chronos: distill historical PRs into cards and typed graphs, retrieve semantic entry plus hops, generate two patches and let a steward select; six SWE backbones. Cut history before the target issue to prevent answer-PR leaks and execute actual patches. |
| 2610.11573 Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies | implementation_or_execution_preconditions | Reviewed MemoType/TriMEM: learned memory/query router selects retrieval under a common Contriever backbone on LongMemEval/LoCoMo/PerLTQA. Train real routing and isolate retrieval labels; manual heuristics do not reproduce it. |
| 2610.11384 Environmental Feedback Modeling Matters: Rethinking Feedback Treatment in Agentic Hindsight Self-Distillation | implementation_or_execution_preconditions | Reviewed SELF: actual on-policy feedback conditions a hindsight teacher; jointly optimize response modeling and self-distillation on reward-free tau/AppWorld training. Need actual gradients for both; privileged future feedback is teacher-only, not deployment. |
| 2610.11226 When Lower Reconstruction Loss Hurts: Distributionally Robust Refinement for Low-Bit LLM Quantization | implementation_or_execution_preconditions | DRQ freezes the grid and refines integer codes against worst-case activation distributions to improve reconstruction generalization. Six PTQ methods on dense/MoE with public PPL/tasks. Require actual integer codes, quantized forwards, and inner uncertainty optimization. |
| 2610.11136 The "10th Juror": Open-Set Standpoint Screening for Bureaucratic Bias Detection | integrity_or_acceptance_audit_preconditions | Reviewed MARS-Gov: legal retrieval, dynamic open-set scout juror, conservative routing, and semantic rewrite checks on DGDB/LOCO. Independently verify evidence and rewrite meaning. Legal context is not advice or unbiasedness assurance. |
| 2610.11123 When Interfaces Speak: Data-Aware Generative UI Harness for Active Interaction | implementation_or_execution_preconditions | Reviewed GenUI tasks: ToolAgent/GUICoder, DynamicUX sandbox interaction rewards, RewardAuditor meta-rubric, real public database simulations in UI-TAU, and 4B training. Require real UI/code/user input and judge-hacking checks. Round questionnaires are not production A/B. |
| 2610.11008 Why LLM Agents Favor Their Group: Stakes, Observed Norms, and Reputation | reference_only | Reviewed group favoritism with arbitrary groups, costless/stakes, scripted histories, and reputation controls over 4,400 simulated societies. Reference for agent behavior only; synthetic point allocation is not organizational causality. |
| 2610.10786 Plan-and-Patch: Diffusion Language Models for Agentic Planning | implementation_or_execution_preconditions | Reviewed Plan-and-Patch: actual diffusion planning via parallel unmasking and regional infill preserving prefix/suffix, with independent executor failure-region selection; NaturalPlan/ALFWorld/TextCraft training. Need dLLM joint generation/repair; autoregressive string editing is not the mechanism. |
| 2610.10305 How Do Transformers Learn to Represent Symmetries? | explicit_unimplemented | Controlled transformer architecture/ablations test group invariance/equivariance in point clouds under finite augmentation. Actual training and OOD geometry are pending, not fixtures. |
| 2610.10120 RewardWeaver: Long-Horizon Interactive Learning for Language Agents via Self-Evolving Reward Adaptation | implementation_or_execution_preconditions | Reviewed RewardWeaver: freeze admitted rubric semantics, attribute low outcomes backward after staged learning, allocate rewards dynamically, and expand under independent controls across social, negotiation, and sales tasks. Require actual training, stable validated rubrics, and no test goals or metric-driven redefinition. |
| 2610.10071 HGP:An on-device personalized agent memory via hybrid graph storage | implementation_or_execution_preconditions | Reviewed HGP: confidence-driven LLM annotations train a lightweight router online; episodic, semantic and procedural graphs plus working state. LoCoMo/PAL selection and QA. Require actual router updates, graph lookup, and on-device time/resources, not label fixtures. |
| 2610.10062 Loud Failures, Quiet Failures: Fault Detection and Recovery in Tool-Using Language Model Agents | implementation_or_execution_preconditions | Reviewed loud/quiet failures: four controlled typed faults injected into a stateful BFCL tool layer, 24 tasks/1,920 trials, fault-free nondeterminism controls, and separate detection/replan/recovery. Real stateful execution; gold final state only scores. Retain reasoning's negative recovery result. |
| 2610.09952 MSU Team at the Explainable Deepfake Detection Challenge 2026: Grounded Artifact Evidence for Deepfake Detection | explicit_unimplemented | XPlainVerse uses DINO/Mesorch detectors, artifact-mask training and patch contrast, plus Qwen-VL explanations and GRPO simplification. Public train/validation/test are available; real multistage implementation pending. |
| 2610.09935 AgentTracer: Tracing Indirect Prompt Injection Attack through Fine-Grained Intention-Execution Alignment | implementation_or_execution_preconditions | Reviewed AgentTracer graph/evaluation: authorize user intent through an operation KB, track implicit task dependencies, and backward-prune from anomalous calls to injection source. Successful attack logs add background noise. Use complete traces and no-attack false-positive controls; do not supply known injection positions. |
| 2610.09856 Self-Evolve With a Reference:Anchored Training of Tool-Integrated Agents | implementation_or_execution_preconditions | Reviewed AnchorLoop: a frozen historical Executor supplies cross-reference advantage and curriculum-consistency proxy, with VeRL sandbox and alternating training. During curriculum the anchor and executor have the same parameters, so consistency is not version gain or correctness. Preserve dual training and snapshots. |
| 2610.09761 PARC-Loc: Text-to-Point-Cloud Localization with Partial Assignment and Relational Consistency | explicit_unimplemented | PARC combines partial assignment, pairwise spatial relations, and neighbor-boundary augmentation on KITTI360/CityLoc. Real text/point-cloud features and coarse/fine weights remain unimplemented. |
| 2610.09624 How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression | implementation_or_execution_preconditions | Reviewed causal tool-call vector: 500 minimal verb pairs, 300 discovery and 200 holdout, hidden-state interventions/transcoder, seven families and native tau2 transfer. Need actual necessity/sufficiency activations, not text heuristics or test-selected vectors. |
| 2610.09484 MIMESIS: Learning User Simulators as Training Environments for Interactive Agents | Full experiment not reproduced; review DONE; original type=implementation_or_execution_preconditions | Meta priority. Reviewed MIMESIS user simulator and agent post-training: user midtraining, ThoughtTrace SFT, multidomain GRPO, freeze simulator, then agent GRPO with CSD token reweighting. Private thoughts are retrospective training-teacher context only. Nine unseen users and eight environments; require both actual training loops, not a fixed simulator. |
| 2610.09430 SkillCycle: Co-Evolving Agent Policies and Skill Banks | explicit_unimplemented | SkillCycle learns skill bank/router under frozen policy and freezes policy for rule repair, using paired rule/bank environment evaluation. Actual ALFWorld/WebShop controller and distillation pending; no-skill deployment cannot use the bank. |
| 2610.09252 Adaptive Visual Token Reduction for Accelerated Image Understanding | explicit_unimplemented | ReFIT reshapes multiscale instruction windows and refines tokens, training reduction modules only. Four VQA FLOP comparisons are not wall-time gains. Actual feature modules pending. |
| 2610.09159 SpecGuard: Proving a Task Is Broken Before the Agent Cheats | Full experiment not reproduced; review DONE; original type=integrity_or_acceptance_audit_preconditions | Google DeepMind priority. Reviewed SpecGuard method, experiments and limits: SpecAgent independently creates Lean specs/test generation is separated in TestAgent. Semantic connector and kernel certificates are conditional on faithful autoformalization. 349 mutated SWE tasks and 60 manually audited specs are not online A/B. Prevent weakened specs and false contradictions. |
| 2610.09127 CADFather: Autonomous CAD Reconstruction through Coordinated Tool Use | implementation_or_execution_preconditions | Reviewed CADFather: VLM reads target/intermediate renders and coordinates pretrained CAD generation and numerical refinement, with actual CadQuery DSL execution and best-valid archive; no new training. Require real 3D geometry and tool policy, and keep test-subset ablations out of parameter selection. |
| 2610.09027 Visual Memory Attacks Can Persist Through The KV Cache | explicit_unimplemented | P-VMI's white-box image influence persists after removal through later KV, tested by cache swap and summary compaction. Real KV interaction and gradient attack remain unimplemented; soft prompts alone are diagnostic. |