Files
ai-taxonomy-literature-anal…/data/labels/o42.jsonl
T
zzyang d6fc7c1dfa Add types-of-work classification, benchmarks and score pages
- Work type = area x role (fixed vs varied), vocabulary in scripts/worktypes.json
- LLM-read labels for 2,678 agent-related papers (data/labels), rule labels for the rest
- Per-type synthesis notes (why-same, sub-variants, benchmarks, workflow)
- Pages: /work-types matrix and detail (timeline, benchmarks, workflow, score charts), /benchmarks
- build_worktypes.py adds work-type tables to data/papers.db

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bhs9LU8eMixk6g1kb1d3j7
2026-10-21 11:55:46 +00:00

61 lines
33 KiB
JSON

{"id": 11726, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "analysis", "why": "Studies effects of memory addition and deletion on agent behavior; finds experience-following, error propagation"}], "pattern": "vary-conditions", "fixes": ["agent frameworks"], "varies": ["memory addition/deletion strategies"], "benchmarks": [], "results": [], "workflow": [], "summary": "Empirical study of how memory addition and deletion affect LLM agents, revealing experience-following, error propagation and misaligned replay."}
{"id": 12576, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "Context rewriting with discourse coherence and response feedback preference data for multi-party dialogue generation"}], "pattern": "vary-data", "fixes": ["dialogue benchmarks"], "varies": ["context rewriting", "preference training"], "benchmarks": [], "results": [], "workflow": ["Rewrite dialogue context", "Build preference data from feedback", "Self-evolve rewriter and responder"], "summary": "DRCR rewrites multi-party dialogue context using discourse coherence and response quality feedback with self-evolution training."}
{"id": 7507, "in_agent_scope": false, "types": [{"area": "train/peft", "role": "method", "why": "Curvature-guided sparse zeroth-order optimization for memory-efficient LLM fine-tuning"}], "pattern": "vary-data", "fixes": ["LLM fine-tuning setup"], "varies": ["ZO parameter selection"], "benchmarks": [], "results": [], "workflow": ["Track curvature from scalar feedback", "Sample coordinates", "Sparse ZO update"], "summary": "CurvZO adaptively selects parameters for sparse zeroth-order fine-tuning using online curvature signals."}
{"id": 4543, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Iterative deep research with Markovian workspace reconstruction and evolving report memory"}, {"area": "search-research", "role": "train-agent", "why": "EAPO RL framework for efficient exploration"}], "pattern": "vary-scaffold", "fixes": ["long-horizon tasks"], "varies": ["context paradigm", "RL reward"], "benchmarks": [], "results": [], "workflow": ["Reconstruct workspace each round", "Synthesize evolving report", "Train with EAPO"], "summary": "IterResearch reformulates deep research as an MDP with workspace reconstruction and trains it with efficiency-aware RL."}
{"id": 10909, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "eval-method", "why": "Hypothesis tests for detecting behavioral change in black-box multi-agent systems via TDKPS"}], "pattern": "vary-evaluation", "fixes": ["multi-agent systems"], "varies": ["detection tests"], "benchmarks": [], "results": [], "workflow": ["Embed agents across time", "Run hypothesis tests"], "summary": "Temporal Data Kernel Perspective Space embeds agents over time with hypothesis tests for detecting behavior shifts."}
{"id": 10153, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Online RL for ML engineering agents with exploration-enriched fine-tuning and step-wise RL"}], "pattern": "fix-env-vary-agent", "fixes": ["ML tasks"], "varies": ["training framework"], "benchmarks": [], "results": [], "workflow": ["Exploration-enriched fine-tuning", "Step-wise RL", "Agentic ML reward"], "summary": "ML-Agent trains LLM agents for autonomous ML engineering with exploration-enriched SFT and step-wise online RL."}
{"id": 11478, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Framework-conditioned four-agent pipeline generating multi-framework front-end code from screenshots"}], "pattern": "vary-scaffold", "fixes": ["MLLM"], "varies": ["agent pipeline"], "benchmarks": [], "results": [], "workflow": ["Ground UI elements", "Plan layout tree", "Write files", "Verify"], "summary": "MulFCoder uses four cooperating agents to generate React/Vue/Angular front-end code from UI screenshots."}
{"id": 626, "in_agent_scope": false, "types": [{"area": "train/data", "role": "method", "why": "Caption trajectory framework TRACED for detecting image-caption errors"}], "pattern": "vary-data", "fixes": ["datasets"], "varies": ["error detection method"], "benchmarks": [], "results": [], "workflow": ["Iteratively edit caption", "Analyze trajectory", "Flag errors"], "summary": "TRACED detects caption errors using trajectories of iterative caption edits that maximize image-text relevance."}
{"id": 14492, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Two-agent pair programming with generator and reviewer role switching for code generation"}], "pattern": "vary-scaffold", "fixes": ["LLM backbones"], "varies": ["collaboration scheme"], "benchmarks": [{"name": "HumanEval", "purpose": "evaluate pass@1 improvement", "is_new": false}], "results": [{"benchmark": "HumanEval", "system": "PairCoder", "base_model": null, "setting": "proposed", "metric": "pass@1", "score": 91.0, "unit": "%"}], "workflow": ["Generator writes code", "Reviewer critiques", "Switch roles on stall"], "summary": "PairCoder pairs two LLMs as coder and reviewer, switching roles when stalled, improving HumanEval with less token cost."}
{"id": 697, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "method", "why": "Predictor of agentic workflow performance using multi-view encoding and pretraining"}], "pattern": "vary-evaluation", "fixes": ["workflows"], "varies": ["performance predictor"], "benchmarks": [], "results": [], "workflow": ["Encode code, prompts, graph", "Pretrain unsupervised", "Predict performance"], "summary": "Agentic Predictor estimates agentic workflow performance via multi-view encoding to cut costly evaluations."}
{"id": 3457, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "SFT then RL for GUI grounding with instruction-as-reasoning"}, {"area": "web-gui", "role": "data", "why": "Cleans flawed grounding instructions and synthesizes diverse ones"}], "pattern": "fix-env-vary-agent", "fixes": ["grounding benchmarks"], "varies": ["instruction perspectives", "training"], "benchmarks": [], "results": [], "workflow": ["Synthesize diverse instructions", "SFT", "RL to select pathway"], "summary": "UI-Ins treats instructions as multi-perspective reasoning pathways, trained with SFT on synthesized instructions then RL."}
{"id": 2801, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "USACOArena credit-budgeted coding arena measuring agent cost decisions"}, {"area": "code-swe", "role": "environment", "why": "Interactive arena with deterministic credit costs"}], "pattern": "fix-agent-vary-env", "fixes": ["agents under test"], "varies": ["budget-constrained tasks"], "benchmarks": [{"name": "USACOArena", "purpose": "cost-aware coding agent evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "USACOArena charges credits for each agent action, exposing cost-benefit decision profiles of coding agents."}
{"id": 1359, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Reranker-guided search over proximity graphs to improve reasoning-intensive retrieval"}], "pattern": "vary-scaffold", "fixes": ["reranker"], "varies": ["retrieval strategy"], "benchmarks": [{"name": "BRIGHT", "purpose": "retrieval gain", "is_new": false}, {"name": "FollowIR", "purpose": "retrieval gain", "is_new": false}, {"name": "M-BEIR", "purpose": "retrieval gain", "is_new": false}], "results": [], "workflow": ["Build proximity graph", "Greedy search by reranker", "Rerank within budget"], "summary": "RGS retrieves documents by reranker preference using greedy graph search, improving BRIGHT, FollowIR and M-BEIR."}
{"id": 10397, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Behavior-invariant task representation with transformer world model for offline meta-RL"}], "pattern": "vary-data", "fixes": ["offline datasets"], "varies": ["representation learning", "conservative penalty"], "benchmarks": [], "results": [], "workflow": ["Learn task latents", "Transformer world model", "Conservative value penalty"], "summary": "Offline meta-RL framework combining information-theoretic task representation with transformer world model."}
{"id": 16000, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Multi-agent progressive filling with global state for document-level scientific information extraction"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["extraction workflow"], "benchmarks": [], "results": [], "workflow": ["Maintain global filling state", "Schedule by priority", "Fill fields progressively"], "summary": "SudokuFill treats scientific information extraction as progressive filling with multi-agent global state."}
{"id": 3811, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Training-free logit averaging across parallel reasoning traces for open-ended reasoning"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["decoding strategy"], "benchmarks": [{"name": "AIME", "purpose": "compare with majority voting", "is_new": false}, {"name": "GPQA", "purpose": "compare with majority voting", "is_new": false}, {"name": "LiveCodeBench", "purpose": "open-ended coding gains", "is_new": false}], "results": [], "workflow": ["Run K parallel traces", "Average logits at sync points"], "summary": "ThinkMerge averages next-token logits from parallel traces, enabling parallel test-time scaling for open-ended reasoning."}
{"id": 7348, "in_agent_scope": false, "types": [{"area": "train/peft", "role": "method", "why": "LLM-guided Bayesian optimization for LoRA hyperparameter search"}], "pattern": "vary-scaffold", "fixes": ["LoRA"], "varies": ["search method"], "benchmarks": [], "results": [], "workflow": ["Prompt LLM with hyperparameter knowledge", "Map to continuous space", "Run BO"], "summary": "Bayesian optimization using LLM-based language-guided mapping to efficiently search LoRA hyperparameters."}
{"id": 16676, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "ConVerse benchmark for privacy and security in agent-to-agent conversations"}, {"area": "multi-agent", "role": "benchmark", "why": "Multi-turn agent-agent negotiation across travel, real estate, insurance"}], "pattern": "fix-agent-vary-env", "fixes": ["agents under test"], "varies": ["attacks", "domains"], "benchmarks": [{"name": "ConVerse", "purpose": "privacy and security risks in agent-agent interactions", "is_new": true}], "results": [], "workflow": [], "summary": "ConVerse benchmarks contextual privacy and security attacks in multi-turn agent-to-agent conversations across three domains."}
{"id": 15995, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Experience-driven early termination reduces SE agent cost"}], "pattern": "vary-scaffold", "fixes": ["SE agents", "model"], "varies": ["termination policy"], "benchmarks": [{"name": "SWE-bench Verified", "purpose": "cost reduction with preserved resolution", "is_new": false}], "results": [], "workflow": ["Extract experience", "Guide early termination", "Select patch"], "summary": "EET uses prior execution experience to terminate SE agents early, cutting cost 19-55% with negligible resolution loss."}
{"id": 7141, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "method", "why": "Communication-conditioned token pruning for V2V cooperative driving MLLMs"}], "pattern": "fix-env-vary-agent", "fixes": ["driving data"], "varies": ["token pruning"], "benchmarks": [], "results": [], "workflow": ["Broadcast request messages", "Identify redundant tokens", "Prune"], "summary": "V2V-CCM prunes visual tokens using request messages to cut communication cost in cooperative driving with MLLMs."}
{"id": 9412, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Tournament graph framework for k-wise zero-shot ranking with query-efficient scheduling"}], "pattern": "vary-scaffold", "fixes": ["oracle"], "varies": ["ranking algorithm"], "benchmarks": [], "results": [], "workflow": ["Aggregate comparisons in graph", "Transitive closure", "Greedy query schedule"], "summary": "BlitzRank uses tournament graphs and information-gain scheduling for efficient top-m ranking with k-wise comparisons."}
{"id": 1232, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "benchmark", "why": "Mix-ECom corpus for mixed-type e-commerce dialogues with domain rules"}], "pattern": "fix-agent-vary-env", "fixes": ["agents under test"], "varies": ["dialogue types", "rules"], "benchmarks": [{"name": "Mix-ECom", "purpose": "evaluate e-commerce agents on mixed dialogues", "is_new": true}], "results": [], "workflow": [], "summary": "Mix-ECom introduces a 4,799-sample e-commerce dialogue corpus with 82 rules for evaluating LLM agents."}
{"id": 17029, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "Train multilingual agentic policy with answer, counter or clarify actions to reduce sycophancy"}], "pattern": "fix-env-vary-agent", "fixes": ["evaluation"], "varies": ["policy training"], "benchmarks": [], "results": [], "workflow": ["Define action space", "Train multi-objective policy"], "summary": "Learns agentic policy controlling sycophancy by choosing to answer, counter misleading signals or ask clarification."}
{"id": 14567, "in_agent_scope": false, "types": [{"area": "train/pretrain", "role": "analysis", "why": "Tests whether data scarcity explains BLiMP failures by injecting 1% synthetic data in GPT-2 pretraining"}], "pattern": "vary-conditions", "fixes": ["GPT-2 Small"], "varies": ["synthetic data injection"], "benchmarks": [{"name": "BLiMP", "purpose": "measure linguistic competence", "is_new": false}], "results": [{"benchmark": "BLiMP only_npi_scope", "system": "GPT-2 Small with injection", "base_model": "GPT-2 Small", "setting": "proposed", "metric": "accuracy", "score": 69.4, "unit": "%"}], "workflow": [], "summary": "Injecting 1% targeted synthetic data into GPT-2 pretraining fixes most worst BLiMP paradigms, implicating data scarcity."}
{"id": 12396, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "eval-method", "why": "Agentic rubrics as execution-free verifiers for SWE patches"}, {"area": "code-swe", "role": "scaffold", "why": "Test-time scaling using rubric scoring"}], "pattern": "vary-scaffold", "fixes": ["SWE agents"], "varies": ["verifier type"], "benchmarks": [{"name": "SWE-Bench Verified", "purpose": "parallel TTS verification", "is_new": false}], "results": [{"benchmark": "SWE-Bench Verified", "system": "Agentic Rubrics", "base_model": "Qwen3-Coder-30B-A3B", "setting": "proposed", "metric": "score", "score": 54.2, "unit": "%"}, {"benchmark": "SWE-Bench Verified", "system": "Agentic Rubrics", "base_model": "Qwen3-32B", "setting": "proposed", "metric": "score", "score": 40.6, "unit": "%"}], "workflow": ["Agent explores repository", "Create rubric checklist", "Score candidate patches"], "summary": "Expert agent builds context-grounded rubrics to score SWE patches without test execution for test-time scaling."}
{"id": 17370, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "GRPO with function calling for dialog state tracking in low-resource settings"}], "pattern": "fix-env-vary-agent", "fixes": ["DST benchmarks"], "varies": ["RL vs SFT training"], "benchmarks": [{"name": "MultiWOZ 2.1", "purpose": "DST evaluation", "is_new": false}, {"name": "MultiWOZ 2.4", "purpose": "DST evaluation", "is_new": false}], "results": [], "workflow": [], "summary": "Applies GRPO with function calling to dialogue state tracking, outperforming SFT with less out-of-domain data."}
{"id": 15822, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "benchmark", "why": "ProMediate testbed for proactive AI mediators in multi-party negotiation"}, {"area": "social-sim", "role": "environment", "why": "Negotiation simulation environment with plug-in mediator"}], "pattern": "fix-agent-vary-env", "fixes": ["mediator agents"], "varies": ["negotiation cases"], "benchmarks": [{"name": "ProMediate", "purpose": "evaluate proactive mediation", "is_new": true}], "results": [], "workflow": [], "summary": "ProMediate simulates multi-party negotiations and offers socio-cognitive metrics to evaluate proactive AI mediators."}
{"id": 15805, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Step-level entropy-triggered experience seeking for web agents"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Experience injection tailored per step"}], "pattern": "vary-scaffold", "fixes": ["Qwen3 models"], "varies": ["experience intervention"], "benchmarks": [], "results": [], "workflow": ["Estimate entropy threshold", "Trigger seeking", "Provide tailored experience"], "summary": "ExpSeek lets web agents seek experience at step-level when entropy is high, improving four benchmarks."}
{"id": 51, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "ChinaTravel open-ended travel planning benchmark with DSL validation"}, {"area": "planning-decision", "role": "environment", "why": "Sandbox for multi-day travel planning"}], "pattern": "fix-agent-vary-env", "fixes": ["language agents"], "varies": ["tasks", "constraints"], "benchmarks": [{"name": "ChinaTravel", "purpose": "travel planning with compositional constraints", "is_new": true}], "results": [], "workflow": [], "summary": "ChinaTravel offers sandbox, DSL evaluation and open-ended dataset for travel planning agents."}
{"id": 12617, "in_agent_scope": false, "types": [{"area": "knowledge/editing", "role": "method", "why": "Causal-guided adaptive layer selection and weight assignment for knowledge editing"}], "pattern": "vary-data", "fixes": ["editing benchmarks"], "varies": ["editing method"], "benchmarks": [], "results": [], "workflow": ["Causal tracing", "Select critical layers", "Allocate edit weights"], "summary": "CAKE selects layers by causal tracing and assigns editing burden through constrained optimization."}
{"id": 18256, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "Evaluates LLMs via semantic parsing into Yarn with SmatchY metric"}, {"area": "eval/bench", "role": "analysis", "why": "Tests inference modes, formats and supervision conditions"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["representation format", "supervision"], "benchmarks": [], "results": [], "workflow": [], "summary": "Evaluates LLM semantic parsing into Yarn under varied settings, showing sensitivity to representation and supervision."}
{"id": 12797, "in_agent_scope": false, "types": [{"area": "knowledge/halluc", "role": "method", "why": "Fine-tuning for answer-dependent verbalized confidence calibration"}], "pattern": "vary-data", "fixes": ["tasks"], "varies": ["fine-tuning method"], "benchmarks": [], "results": [], "workflow": ["Diagnose answer-independence", "Fine-tune answer-grounded confidence"], "summary": "ADVICE fine-tunes LLMs to condition verbalized confidence on their own answers, improving calibration."}
{"id": 14973, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Hybrid memory with graph, experience and passage components for long-horizon agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory design"], "benchmarks": [], "results": [], "workflow": ["Build graph memory", "Abstract experience memory", "Dual-channel retrieval"], "summary": "MemWeaver unifies temporal graph, experience and passage memories with dual-channel retrieval."}
{"id": 13783, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "Two-stage optimization learning memory organization guideline and RL update policy"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["memory update policy"], "benchmarks": [], "results": [], "workflow": ["Induce memory guideline", "Train memory policy with RL"], "summary": "MemCoE learns how to organize and what to update in user memory via guideline induction and policy optimization."}
{"id": 15220, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "analysis", "why": "Finds TIR models fail to trust tool results"}, {"area": "tool-use", "role": "train-agent", "why": "Adaptive tool trust calibration for TIR math reasoning"}], "pattern": "fix-env-vary-agent", "fixes": ["TIR models"], "varies": ["tool trust training"], "benchmarks": [], "results": [], "workflow": [], "summary": "Analyzes tool-integrated reasoning models ignoring correct tool outputs and proposes adaptive tool trust calibration."}
{"id": 8176, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "benchmark", "why": "HOI-Edit benchmark with HOI-Eval metric for human-object interaction editing"}, {"area": "arch/multimodal", "role": "scaffold", "why": "Agentic self-correcting framework using I2V models"}], "pattern": "fix-agent-vary-env", "fixes": ["I2V models"], "varies": ["editing tasks"], "benchmarks": [{"name": "HOI-Edit", "purpose": "human-object interaction image editing", "is_new": true}], "results": [], "workflow": [], "summary": "HOI-Edit benchmarks interaction editing and uses I2V models with agentic self-correction."}
{"id": 16041, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "benchmark", "why": "MAS-BENCH distributed sorting benchmark exposing coordination failures"}, {"area": "multi-agent", "role": "scaffold", "why": "CAMOC collaboration-aware information sharing protocol"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM agents"], "varies": ["agent count", "communication mode"], "benchmarks": [{"name": "MAS-BENCH", "purpose": "multi-agent coordination under communication constraints", "is_new": true}], "results": [], "workflow": ["Share early metadata", "Single-commit verification"], "summary": "MAS-BENCH shows LLM multi-agent coordination collapses with scale; CAMOC mitigates it."}
{"id": 16355, "in_agent_scope": false, "types": [{"area": "multiling/mt", "role": "eval-method", "why": "LQM linguistically motivated error taxonomy for Arabic dialect MT"}, {"area": "multiling/mt", "role": "benchmark", "why": "Parallel corpus of 3,850 sentences across seven dialects"}], "pattern": "vary-evaluation", "fixes": ["MT systems"], "varies": ["metric taxonomy"], "benchmarks": [{"name": "LQM corpus", "purpose": "Arabic dialect MT evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "LQM provides hierarchical linguistic error taxonomy and dialect corpus for evaluating Arabic machine translation."}
{"id": 12933, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Tool-augmented multi-agent zoom-in for long visual document understanding"}], "pattern": "vary-scaffold", "fixes": ["Gemini-2.5-Pro"], "varies": ["agent framework"], "benchmarks": [{"name": "MMLongBench-Doc", "purpose": "state-of-the-art evaluation", "is_new": false}, {"name": "FinRAGBench-V", "purpose": "state-of-the-art evaluation", "is_new": false}], "results": [], "workflow": ["Navigate to relevant pages", "Zoom into elements", "Sampling-adjudication"], "summary": "DocLens zooms into evidence with tool-augmented agents for long visual documents, achieving state-of-the-art."}
{"id": 14909, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "AP-Test identifies guardrails in black-box agents with adversarial prompts"}], "pattern": "vary-conditions", "fixes": ["black-box agent"], "varies": ["adversarial prompts"], "benchmarks": [], "results": [], "workflow": ["Input guard test", "Output guard test", "Match score"], "summary": "AP-Test uses guard-specific adversarial prompts to identify deployed guardrails in black-box agents."}
{"id": 2233, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Lifelong embodied navigation with DE-LoRA decoupled shared and task-specific knowledge"}], "pattern": "fix-env-vary-agent", "fixes": ["task sequence"], "varies": ["lifelong learning method"], "benchmarks": [], "results": [], "workflow": ["Shared knowledge inheritance", "Experts co-activation", "Task-specific learning"], "summary": "Uni-Walker uses DE-LoRA to continually learn navigation tasks without catastrophic forgetting."}
{"id": 14034, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "PaperMentor multi-agent writing tutor for Overleaf giving inline suggestions"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["domain pipeline"], "benchmarks": [], "results": [], "workflow": ["Skill library", "12 specialized agents", "Overleaf inline comments"], "summary": "PaperMentor is a multi-agent Overleaf assistant giving actionable writing feedback on research papers."}
{"id": 8407, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "LLM agent with ReAct loop for in silico protein design via directed mutation"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["domain pipeline"], "benchmarks": [], "results": [], "workflow": ["Think and plan", "Act with tools", "Observe and reflect"], "summary": "PDAgent is a ReAct-style autonomous agent for template-based directed protein mutation."}
{"id": 16215, "in_agent_scope": false, "types": [{"area": "reason/code", "role": "method", "why": "RL training internalizing code generation, reflection and self-correction"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["RL reward"], "benchmarks": [], "results": [], "workflow": ["Generate", "Reflect", "Self-correct", "RL"], "summary": "ReflexiCoder uses RL to internalize reflection and self-correction for code generation."}
{"id": 15460, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "SciExplore benchmark of 103 scientific information-seeking tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs and agents"], "varies": ["task types"], "benchmarks": [{"name": "SciExplore", "purpose": "scientific information-seeking evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "SciExplore evaluates agents on scientific database navigation, literature retrieval and knowledge synthesis."}
{"id": 5371, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Multi-agent LLM search for ODEs with qualitative and quantitative evaluation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["discovery workflow"], "benchmarks": [], "results": [], "workflow": ["Sampler proposes candidates", "Optimize parameters", "Scientist agent evaluates"], "summary": "DoLQ discovers ODEs with sampler, optimizer and scientist agents combining qualitative and quantitative evaluation."}
{"id": 15365, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "LifeSim BDI user simulator for long-horizon life trajectories"}, {"area": "social-sim", "role": "benchmark", "why": "LifeSim-Eval personalized assistance benchmark, 1,200 scenarios"}], "pattern": "fix-agent-vary-env", "fixes": ["assistants"], "varies": ["simulated scenarios"], "benchmarks": [{"name": "LifeSim-Eval", "purpose": "personalized assistant evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "LifeSim simulates users with BDI cognition and introduces LifeSim-Eval for long-horizon personalized assistants."}
{"id": 13253, "in_agent_scope": false, "types": [{"area": "interp/mech", "role": "analysis", "why": "Shows multilingual language units depend on script via LAPE and sparse autoencoders"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["romanization", "word order"], "benchmarks": [], "results": [], "workflow": [], "summary": "Multilingual LMs encode orthographic script more than linguistic structure."}
{"id": 287, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Multi-scale autoregressive generation for long-horizon offline RL"}], "pattern": "fix-env-vary-agent", "fixes": ["offline RL benchmarks"], "varies": ["generative method"], "benchmarks": [], "results": [], "workflow": ["Multi-scale autoencoder", "Multi-scale transformer"], "summary": "MAGE generates trajectories via multi-scale autoregression for sparse-reward long-horizon offline RL."}
{"id": 17310, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "application", "why": "elfen Python library for linguistic feature extraction"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "elfen is a Python package extracting item-level linguistic features across eleven areas for text datasets."}
{"id": 15932, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "method", "why": "Heterogeneous graph contrastive learning router for selecting agents and collaboration modes"}], "pattern": "vary-scaffold", "fixes": ["agents"], "varies": ["routing"], "benchmarks": [], "results": [], "workflow": ["Build heterogeneous graph", "Predict edge weights", "Contrastive learning"], "summary": "RouterHGC routes queries to agent roles, collaboration modes and LLMs via heterogeneous graph contrastive learning."}
{"id": 10949, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "theory", "why": "Probabilistic theory of agency with logarithmic pooling and Waluigi alignment formalization"}], "pattern": "n/a", "fixes": [], "varies": ["assumptions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Develops theory of agents as outcome distributions and formalizes persona misalignment in LLMs."}
{"id": 14581, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "VLM rationale-guided logic reasoning for multimodal misinformation detection"}], "pattern": "vary-scaffold", "fixes": ["VLM"], "varies": ["detection framework"], "benchmarks": [], "results": [], "workflow": ["Generate verification rationale", "Guide feature fusion"], "summary": "LoGAR uses VLM-generated verification rationales to condition multimodal misinformation detection."}
{"id": 12752, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Multi-armed bandit for aspect-aware retrieval in multimodal document QA"}], "pattern": "vary-scaffold", "fixes": ["retriever"], "varies": ["aspect weighting"], "benchmarks": [], "results": [], "workflow": ["Decompose query to subqueries", "Retrieve per aspect", "Bandit selection"], "summary": "MAB-DQA models query aspect importance with bandits for multimodal document question answering."}
{"id": 16515, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "survey", "why": "Position paper on responsible LLM use in peer review"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Argues LLMs should support reproducibility, citation checks and ethics flagging in peer review, not generate reviews."}
{"id": 12568, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "LLM label induction and contrastive text-label alignment for generalized category discovery"}], "pattern": "vary-data", "fixes": ["datasets"], "varies": ["alignment method"], "benchmarks": [], "results": [], "workflow": ["Dual-encoder contrastive", "LLM label induction", "Graph refinement"], "summary": "TLSA aligns text and label spaces and uses LLM-induced labels for generalized category discovery."}
{"id": 15394, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "defense", "why": "Ontological constraints enforcing action admissibility for LLM agents"}], "pattern": "vary-scaffold", "fixes": ["LLM agent"], "varies": ["admissibility guard"], "benchmarks": [], "results": [], "workflow": ["Extract constraints from demonstrations", "Enforce at inference"], "summary": "OntoGuard enforces environmental admissibility on LLM agent actions using ontology from demonstrations and world knowledge."}
{"id": 13641, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "survey", "why": "Survey of 58 foundation and agentic models for single-cell biology"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "LLM4Cell surveys 58 foundation and agentic models for single-cell biology with taxonomy and evaluation."}
{"id": 9505, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "eval-method", "why": "SWE-ABS strengthens test suites to expose incorrect passing patches"}], "pattern": "vary-evaluation", "fixes": ["agents' patches"], "varies": ["test suites"], "benchmarks": [{"name": "SWE-Bench Verified", "purpose": "re-evaluation with strengthened tests", "is_new": false}], "results": [], "workflow": ["Coverage-driven augmentation", "Mutation adversarial testing"], "summary": "SWE-ABS strengthens SWE-Bench Verified tests, rejecting 19.78% of passing patches and lowering top score."}
{"id": 15942, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Two-stage multi-agent LLM framework for popular path query"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent pipeline"], "benchmarks": [], "results": [], "workflow": ["SEARCH stage", "Generate popular path"], "summary": "CompassLLM applies multi-agent LLMs to geo-spatial popular path queries without retraining."}