d6fc7c1dfa
- 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
61 lines
31 KiB
JSON
61 lines
31 KiB
JSON
{"id": 12283, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "Cost-centric travel-planning benchmark with dynamic blocking events for tool-use agents"}, {"area": "planning-decision", "role": "analysis", "why": "Shows gap in cost-aware planning and replanning across models"}], "pattern": "fix-agent-vary-env", "fixes": ["models under test"], "varies": ["tool costs", "blocking events", "tasks"], "benchmarks": [{"name": "CostBench", "purpose": "evaluate cost-optimal planning", "is_new": true}], "results": [], "workflow": [], "summary": "CostBench evaluates cost-optimal planning and adaptation of tool-use agents under dynamic blocking events in travel planning."}
|
|
{"id": 17589, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Multiple-choice multi-hop code comprehension dataset from SWE-bench repositories"}], "pattern": "fix-agent-vary-env", "fixes": ["models under test"], "varies": ["tasks"], "benchmarks": [{"name": "SWE-QA", "purpose": "multi-hop code comprehension", "is_new": true}, {"name": "SWE-bench", "purpose": "source repositories", "is_new": false}], "results": [], "workflow": [], "summary": "SWE-QA is a 9,072-question multiple-choice benchmark for multi-hop code understanding built from SWE-bench repositories."}
|
|
{"id": 15194, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Two-stage RL using post-editing auxiliary task to stabilize MT training"}], "pattern": "vary-conditions", "fixes": ["benchmarks"], "varies": ["RL method"], "benchmarks": [], "results": [], "workflow": ["Sample translations", "Build post-editing inputs", "Joint two-stage RL"], "summary": "PEGRL uses post-editing as auxiliary task in RL to lower variance and improve LLM machine translation."}
|
|
{"id": 14247, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Runtime interception framework validating MCP tool calls and responses"}, {"area": "agent-safety", "role": "survey", "why": "Threat taxonomy from 80+ MCP attack techniques"}], "pattern": "n/a", "fixes": ["threat model"], "varies": ["protective mechanism"], "benchmarks": [], "results": [], "workflow": ["Intercept MCP calls", "Structural analysis", "Semantic intent verification"], "summary": "ShieldMCP provides an MCP threat taxonomy and runtime defense that validates tool calls and responses."}
|
|
{"id": 15463, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Training-free iterative query refinement using retrieved candidates for zero-shot composed image retrieval"}], "pattern": "vary-conditions", "fixes": ["benchmarks"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": ["Retrieve candidates", "MLLM inspects misalignment", "Fuse historical queries"], "summary": "CoRR refines queries with retrieval feedback and MLLM CoT for zero-shot composed image retrieval."}
|
|
{"id": 13058, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Entropy-guided adaptive test-time scaling with test-suite augmentation for SWE agents"}], "pattern": "vary-scaffold", "fixes": ["model weights", "environment"], "varies": ["test-time scaling procedure"], "benchmarks": [{"name": "SWE-bench Verified", "purpose": "evaluate TTS gains", "is_new": false}], "results": [], "workflow": ["Entropy-guided stepwise search", "Augment test suites", "Select best candidate"], "summary": "EGSS adaptively scales test-time compute for software engineering agents using entropy and augmented tests."}
|
|
{"id": 15164, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Collaborative agents jointly optimizing prompts and decoding hyperparameters"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["prompt and decoding config"], "benchmarks": [], "results": [], "workflow": ["Encode prompt+hyperparameters as agents", "Collaborative search", "Iterate"], "summary": "Agent-GWO optimizes prompt templates and decoding hyperparameters jointly via collaborative agents for reasoning."}
|
|
{"id": 16094, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "data", "why": "Multi-agent pipeline synthesizes EMS dialogues from ePCR for diagnosis prediction training"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training data"], "benchmarks": [{"name": "EMSDialog", "purpose": "synthetic EMS conversation dataset", "is_new": true}], "results": [], "workflow": ["Plan topic flow", "Generate dialogue", "Self-refine with checks"], "summary": "EMSDialog is a synthetic multi-speaker EMS conversation dataset generated by multi-LLM agents from ePCRs."}
|
|
{"id": 5619, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Reflective LLM-agent meta-optimization for analog circuit transistor sizing"}], "pattern": "vary-scaffold", "fixes": ["general LLM"], "varies": ["sizing pipeline"], "benchmarks": [], "results": [], "workflow": ["Inner loop circuit sizing", "Outer loop analyzes optimization", "Adapt search space"], "summary": "AutoSizer is a two-loop LLM agent framework for automatic sizing of analog and mixed-signal circuits."}
|
|
{"id": 18427, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "benchmark", "why": "AMR corpus of classroom language for pedagogical conversational agents"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [{"name": "JIA-AMRs Collection", "purpose": "AMR resource for classroom language", "is_new": true}], "results": [], "workflow": [], "summary": "JIA-AMRs is a corpus of Abstract Meaning Representations of classroom student and curricular language."}
|
|
{"id": 514, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Retrieval-augmented world model grounding LLM simulation for computer-use agents"}, {"area": "web-gui", "role": "analysis", "why": "Probes LLM world-model ability on next-state, planning and milestone tasks"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["world model retrieval"], "benchmarks": [], "results": [], "workflow": ["Retrieve grounding knowledge", "Simulate future states", "Estimate reward"], "summary": "R-WoM augments LLM world models with retrieved knowledge to reduce compounding errors for computer-use agents."}
|
|
{"id": 15529, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "One-step physics-consistent flow policy with elastic time horizons for manipulation"}], "pattern": "vary-conditions", "fixes": ["benchmarks"], "varies": ["policy method"], "benchmarks": [{"name": "LIBERO", "purpose": "manipulation evaluation", "is_new": false}, {"name": "CALVIN", "purpose": "manipulation evaluation", "is_new": false}, {"name": "RoboTwin", "purpose": "manipulation evaluation", "is_new": false}], "results": [], "workflow": [], "summary": "ElasticFlow is a distillation-free one-step policy for fast language-guided robotic manipulation."}
|
|
{"id": 1058, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Entity-aware GRPO rewarding near-miss samples using ground-truth entities for search agents"}], "pattern": "fix-env-vary-agent", "fixes": ["synthetic data", "benchmarks"], "varies": ["reward and RL algorithm"], "benchmarks": [], "results": [], "workflow": ["Count identified entities", "Entity-aware reward", "Train with E-GRPO"], "summary": "E-GRPO adds entity-based dense rewards to GRPO so search agents learn from near-miss trajectories."}
|
|
{"id": 18147, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Graph attention model predicting claim impact from argument graphs"}], "pattern": "vary-conditions", "fixes": ["Kialo data"], "varies": ["model architecture"], "benchmarks": [{"name": "Kialo", "purpose": "debate data for training", "is_new": false}], "results": [], "workflow": ["Encode claims", "Graph attention over relations", "Predict impact distribution"], "summary": "A GAT model on argument graphs predicts claim impact vote distributions on Kialo debates."}
|
|
{"id": 18376, "in_agent_scope": false, "types": [{"area": "interp/mech", "role": "analysis", "why": "Circuit analysis of locative preposition selection in LLaMA-3.2-1b"}], "pattern": "vary-conditions", "fixes": ["model"], "varies": ["circuit interventions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Mechanistic interpretability study finding circuits for locative preposition choice in LLaMA-3.2-1b."}
|
|
{"id": 8379, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Hierarchical view and token selection for zero-shot 3D question answering with VLMs"}], "pattern": "vary-conditions", "fixes": ["VLMs"], "varies": ["input view/token selection"], "benchmarks": [], "results": [], "workflow": ["Score view importance", "Select views", "Transport-based token selection"], "summary": "KeyVT selects views and representative tokens for zero-shot 3D QA with 2D VLMs."}
|
|
{"id": 13242, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Learns finite-state transition and communication weights for multi-agent coordination"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["coordination policy"], "benchmarks": [], "results": [], "workflow": ["Model as finite-state process", "Learn transitions from traces", "Trust-aware message routing"], "summary": "NeuralFSM learns state transitions and inter-agent communication for adaptive, robust multi-agent coordination."}
|
|
{"id": 8550, "in_agent_scope": true, "types": [{"area": "search-research", "role": "eval-method", "why": "Battle-style expert evaluation protocol for literature review quality"}, {"area": "search-research", "role": "benchmark", "why": "Collects 3k expert judgments comparing review systems"}], "pattern": "vary-evaluation", "fixes": ["systems under test"], "varies": ["evaluation protocol"], "benchmarks": [{"name": "LitReview Arena", "purpose": "expert comparison of literature reviews", "is_new": true}], "results": [{"benchmark": "LitReview Arena", "system": "strongest systems vs human drafts", "base_model": null, "setting": "baseline", "metric": "win rate of decisive matches", "score": 23.0, "unit": "%"}], "workflow": [], "summary": "LitReview Arena is an expert pairwise platform evaluating literature review agents across five criteria."}
|
|
{"id": 5061, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Evidence-grounded agentic framework combining VLM and segmentation model for medical VQA"}, {"area": "domain-agents", "role": "train-agent", "why": "RL with verifiable rewards for the compact VLM"}], "pattern": "vary-scaffold", "fixes": ["general models"], "varies": ["agentic pipeline"], "benchmarks": [], "results": [], "workflow": ["VLM proposes entities", "Segmentation yields ROIs", "Reason over evidence"], "summary": "CARE decomposes medical reasoning into grounded sub-modules for accountable multimodal medical VQA."}
|
|
{"id": 13935, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "train-agent", "why": "Learns mapping from flawed RAG outputs to corrective action plans without taxonomies"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training paradigm"], "benchmarks": [], "results": [], "workflow": ["Collect flawed outputs", "Map to action plans", "Train response-action model"], "summary": "RePAIR trains RAG correction directly from flawed outputs to action plans without error taxonomies."}
|
|
{"id": 15059, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "61,845 malicious URL attack instances for web agents"}, {"area": "agent-safety", "role": "defense", "why": "URLGuard lightweight defense module"}], "pattern": "fix-agent-vary-env", "fixes": ["12 LLMs"], "varies": ["URL attack scenarios"], "benchmarks": [{"name": "MalURLBench", "purpose": "malicious URL vulnerability", "is_new": true}], "results": [], "workflow": [], "summary": "MalURLBench benchmarks LLM vulnerabilities to disguised malicious URLs and proposes URLGuard defense."}
|
|
{"id": 12799, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Selective branching at critical states for resource-efficient agentic RL exploration"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["rollout exploration strategy"], "benchmarks": [], "results": [], "workflow": ["Identify key states", "Branch dynamically", "Allocate rollouts"], "summary": "SPARK branches rollouts at critical decision states for efficient long-horizon agent RL."}
|
|
{"id": 17418, "in_agent_scope": false, "types": [{"area": "safety/fairness", "role": "analysis", "why": "Audit of skin-tone bias in text-to-image generation under income framing"}], "pattern": "vary-conditions", "fixes": ["T2I models"], "varies": ["income wording in prompts"], "benchmarks": [{"name": "MST-E", "purpose": "skin tone annotation benchmark", "is_new": false}], "results": [], "workflow": [], "summary": "Audit showing income and occupation cues in prompts shift skin tones in text-to-image portraits."}
|
|
{"id": 14376, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Confidence-guided contrastive decoding over decomposed premises"}], "pattern": "vary-conditions", "fixes": ["model"], "varies": ["decoding method"], "benchmarks": [], "results": [], "workflow": ["Decompose premises", "Estimate confidence", "Contrast logits"], "summary": "C^3D contrasts logits from high and low confidence premises to reduce distraction during reasoning."}
|
|
{"id": 10198, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "Agents rediscover verified findings from ML papers given research questions"}], "pattern": "fix-agent-vary-env", "fixes": ["frontier agents"], "varies": ["research tasks"], "benchmarks": [{"name": "FIRE-Bench", "purpose": "scientific insight rediscovery", "is_new": true}], "results": [], "workflow": [], "summary": "FIRE-Bench evaluates agents on rediscovering established ML research findings with verifiable scoring."}
|
|
{"id": 12073, "in_agent_scope": true, "types": [{"area": "search-research", "role": "survey", "why": "Systematic survey of LLM-based search agents"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "First systematic survey of search agents covering architecture, optimization, application and evaluation."}
|
|
{"id": 11125, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "1,865 long-horizon software engineering problems across 41 repositories"}], "pattern": "fix-agent-vary-env", "fixes": ["agents under test"], "varies": ["tasks"], "benchmarks": [{"name": "SWE-Bench Pro", "purpose": "long-horizon SWE evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "SWE-Bench Pro is a benchmark of complex professional software engineering tasks with public and private sets."}
|
|
{"id": 8207, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Belief-change intrinsic reward for credit assignment in information-seeking RL"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["reward signal"], "benchmarks": [], "results": [], "workflow": ["Measure belief change", "Reward progress", "RL on synthetic interactions"], "summary": "DeltaBelief-RL rewards intermediate progress using change in the model's belief about the target solution."}
|
|
{"id": 6092, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Multi-agent debate with long-term memory to mitigate unsafe behavior"}, {"area": "agent-safety", "role": "attack", "why": "Automated red-teaming through debate uncovers failure modes"}], "pattern": "vary-scaffold", "fixes": ["models"], "varies": ["debate and memory"], "benchmarks": [], "results": [], "workflow": ["Multi-LLM debate", "Find unsafe failures", "Store insights in memory"], "summary": "RedDebate uses multi-agent red-teaming debates and memory modules to make LLM responses safer."}
|
|
{"id": 11631, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Spectral-decoupled iterative refinement for precipitation nowcasting"}], "pattern": "vary-conditions", "fixes": ["benchmarks"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": ["Extract low-frequency skeleton", "Refine high frequency", "Physical constraints"], "summary": "SDIR is a deterministic frequency-decoupled refinement framework for precipitation nowcasting."}
|
|
{"id": 18235, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "data", "why": "Detection-guided LLM pipeline building Japanese caption and VQA datasets"}], "pattern": "vary-data", "fixes": [], "varies": ["dataset"], "benchmarks": [{"name": "DEJIMA-Cap", "purpose": "Japanese image captions", "is_new": true}, {"name": "DEJIMA-VQA", "purpose": "Japanese visual QA", "is_new": true}], "results": [], "workflow": ["Detect objects", "LLM generates Japanese text", "Build datasets"], "summary": "DEJIMA builds large Japanese image-caption and VQA datasets via detection-guided LLM refinement."}
|
|
{"id": 17482, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "benchmark", "why": "Persona-guided shopping benchmark built on WebShop and tau2-Bench"}, {"area": "social-sim", "role": "eval-method", "why": "New persona fidelity and alignment metrics"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["personas", "tasks"], "benchmarks": [{"name": "ShopperBench", "purpose": "personalized shopping", "is_new": true}, {"name": "WebShop", "purpose": "base dataset", "is_new": false}, {"name": "tau2-Bench", "purpose": "base environment", "is_new": false}], "results": [], "workflow": [], "summary": "ShopperBench evaluates persona-adaptive shopping agents with persona-conditioned goals and new metrics."}
|
|
{"id": 10594, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Bilevel graph-based planning and tool execution for scalable tool orchestration"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["planning architecture"], "benchmarks": [], "results": [], "workflow": ["Plan response or toolchain", "Tool graph navigation", "Execute toolchain"], "summary": "NaviAgent decouples planning from tool execution using a tool-relation graph model for large tool sets."}
|
|
{"id": 9923, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Hindsight policy optimization with intent-space Wasserstein signals for long-horizon agents"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["policy gradient method"], "benchmarks": [], "results": [], "workflow": ["Project to intent space", "Wasserstein distance signal", "Policy update"], "summary": "HPO extracts low-variance credit signals from hindsight distributions for long-horizon language agent RL."}
|
|
{"id": 1305, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Two-stage training teaching agents to simulate futures via search-tree reasoning traces"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["training framework"], "benchmarks": [], "results": [], "workflow": ["Build trees from experience", "Train ReSim reasoning", "Second-stage training"], "summary": "Dyna-Mind trains agents to mentally simulate alternatives for web and computer use tasks."}
|
|
{"id": 857, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "Asynchronous dynamic-environment agent benchmark with action-level verifier"}, {"area": "multi-agent", "role": "benchmark", "why": "Includes collaboration with other agents"}], "pattern": "fix-agent-vary-env", "fixes": ["models under test"], "varies": ["scenarios"], "benchmarks": [{"name": "Gaia2", "purpose": "asynchronous agent evaluation", "is_new": true}], "results": [{"benchmark": "Gaia2", "system": "GPT-5 (high)", "base_model": null, "setting": "baseline", "metric": "pass@1", "score": 42, "unit": "%"}], "workflow": [], "summary": "Gaia2 benchmarks agents in evolving asynchronous environments with write-action verifiers usable for RLVR."}
|
|
{"id": 18169, "in_agent_scope": false, "types": [{"area": "multiling/multi", "role": "analysis", "why": "Human study of factors affecting translation of Polish noun phrases"}], "pattern": "vary-conditions", "fixes": [], "varies": ["word order", "international words"], "benchmarks": [], "results": [], "workflow": [], "summary": "Online human experiment on linguistic and demographic factors in free translation of Polish phrases."}
|
|
{"id": 18380, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "analysis", "why": "Cognitive-load linguistic cues predicting listener understanding in explanations"}], "pattern": "vary-conditions", "fixes": [], "varies": ["linguistic cues"], "benchmarks": [{"name": "MUNDEX", "purpose": "dialogue corpus", "is_new": false}], "results": [], "workflow": [], "summary": "Statistical analysis of surprisal, complexity and gaze cues versus listener understanding states."}
|
|
{"id": 209, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "Evolves existing agent tasks into harder ones with recorded trajectories"}, {"area": "planning-decision", "role": "eval-method", "why": "Self-evolving benchmark framework"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["task difficulty"], "benchmarks": [{"name": "TRACE", "purpose": "benchmark evolution framework", "is_new": true}], "results": [], "workflow": ["Mine evolution proposals", "Form problems with exploration", "Validate trajectories"], "summary": "TRACE lets agents evolve existing benchmark tasks into harder, verifiable ones with traceable trajectories."}
|
|
{"id": 13164, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Interactive decision framework with Bayesian preference elicitation from documents"}], "pattern": "vary-scaffold", "fixes": ["general LLM"], "varies": ["decision pipeline"], "benchmarks": [], "results": [], "workflow": ["Extract option matrix", "Adaptive pairwise questions", "Bayesian preference inference"], "summary": "Decisive grounds decisions in document-derived scores and elicits user preferences adaptively."}
|
|
{"id": 10139, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "analysis", "why": "Causal interventions testing experience faithfulness across 13 backbones and 9 environments"}], "pattern": "vary-conditions", "fixes": ["frameworks"], "varies": ["experience form"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows self-evolving agents rely on raw but often ignore condensed experience."}
|
|
{"id": 11355, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "CAPTCHA framework exploiting human-agent cognitive gap"}, {"area": "agent-safety", "role": "benchmark", "why": "Scalable CAPTCHA generation benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["CAPTCHA types"], "benchmarks": [{"name": "Next-Gen CAPTCHAs", "purpose": "agent defense benchmark", "is_new": true}, {"name": "OpenCaptchaWorld", "purpose": "prior baseline", "is_new": false}], "results": [], "workflow": [], "summary": "Scalable CAPTCHA generation framework resisting advanced GUI agents."}
|
|
{"id": 16344, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Entropy-weighted advantage normalization for multi-answer RLVR"}], "pattern": "vary-conditions", "fixes": ["tasks"], "varies": ["advantage normalization"], "benchmarks": [], "results": [], "workflow": ["Compute token entropy", "Weight reward statistics", "Decay schedule"], "summary": "EARS reshapes RLVR advantages with entropy for multi-answer reasoning."}
|
|
{"id": 4918, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "Two-stage visual token pruning from Mamba parameters"}], "pattern": "vary-conditions", "fixes": ["benchmarks"], "varies": ["pruning method"], "benchmarks": [], "results": [], "workflow": ["Score tokens by delta", "Early pruning", "Late complete pruning"], "summary": "DTP prunes visual tokens in Mamba-based MLLMs, cutting computation by nearly 50%."}
|
|
{"id": 13715, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Gradient-free evolutionary optimization of modular tool-use policy with blame attribution"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["module prompts"], "benchmarks": [], "results": [], "workflow": ["Blame attribution", "Targeted mutation", "Diversity-aware selection"], "summary": "EvoTool evolves a modular Planner/Selector/Caller/Synthesizer tool-use policy without gradients."}
|
|
{"id": 14134, "in_agent_scope": false, "types": [{"area": "multiling/multi", "role": "analysis", "why": "Fine-tunes seven LLMs on Arabic testing Semitic-specific transfer"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["language family", "CoT"], "benchmarks": [], "results": [], "workflow": [], "summary": "Finds no Semitic-specific transfer; gains reflect task-format alignment."}
|
|
{"id": 4968, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "analysis", "why": "Systematic study of LLM poker play and reasoning flaws"}, {"area": "tool-use", "role": "scaffold", "why": "ToolPoker tool-integrated reasoning for game-theoretic play"}], "pattern": "vary-conditions", "fixes": ["games"], "varies": ["models", "tool integration"], "benchmarks": [], "results": [], "workflow": ["Call solver tools", "Integrate with reasoning"], "summary": "Analyzes LLM poker failures and proposes ToolPoker tool-integrated reasoning."}
|
|
{"id": 17154, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Language-grounded multi-domain image translation with semantic difference guidance"}], "pattern": "vary-conditions", "fixes": ["benchmarks"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": ["GLIP-Adapter fusion", "Ground prompt delta", "Per-attribute control"], "summary": "LACE enables attribute-controllable multi-domain image translation from language differences."}
|
|
{"id": 18600, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "analysis", "why": "Tracks idiomaticity preference during Swedish pretraining and adaptation"}, {"area": "eval/bench", "role": "benchmark", "why": "Two new Swedish idiom datasets"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["training checkpoints"], "benchmarks": [], "results": [], "workflow": [], "summary": "Idiomatic competence emerges slowly in Swedish models; two new minimal-pair idiom datasets introduced."}
|
|
{"id": 15999, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Multi-constraint cross-dimensional spatial reasoning for geometric visual questions"}], "pattern": "vary-conditions", "fixes": ["task"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Method for cross-dimensional geometric visual QA using 2D relations and 3D topology."}
|
|
{"id": 14265, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Training-free schema-conditioned agentic reasoning over enterprise knowledge graphs"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["reasoning framework"], "benchmarks": [{"name": "CMDB benchmark", "purpose": "enterprise KG evaluation", "is_new": false}], "results": [], "workflow": ["Inject schema priors", "Plan structurally", "Schema-aware traversal"], "summary": "SCAIR is a training-free schema-aware agentic KG-RAG framework for enterprise knowledge graphs."}
|
|
{"id": 12306, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "AST-entity structured action space with readCode and editCode for code agents"}], "pattern": "vary-scaffold", "fixes": ["models"], "varies": ["action interface"], "benchmarks": [{"name": "SWE-Bench Verified", "purpose": "code agent evaluation", "is_new": false}, {"name": "CodeAssistBench", "purpose": "code agent evaluation", "is_new": false}], "results": [], "workflow": [], "summary": "CodeStruct gives code agents syntax-aware structured actions, improving accuracy and token cost."}
|
|
{"id": 6762, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "train-agent", "why": "Pattern-aware alignment and balanced reward for time series QA"}], "pattern": "vary-conditions", "fixes": ["tasks"], "varies": ["model and reward"], "benchmarks": [], "results": [], "workflow": [], "summary": "PATRA aligns time series patterns and balances rewards for time series reasoning."}
|
|
{"id": 17285, "in_agent_scope": false, "types": [{"area": "align/pref", "role": "analysis", "why": "Shows process reward models fail on reasoning models, using probes and SAEs"}], "pattern": "vary-conditions", "fixes": ["PRMs"], "varies": ["model type"], "benchmarks": [], "results": [], "workflow": [], "summary": "Analysis finding PRMs degrade reasoning models and respond to formatting artifacts."}
|
|
{"id": 16929, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "defense", "why": "Pretrained guide models with contrastive fine-tuning for LLM domain certification"}], "pattern": "vary-conditions", "fixes": ["VALID certification"], "varies": ["guide model"], "benchmarks": [], "results": [], "workflow": [], "summary": "PRISM uses pretrained guide models to improve domain certification over VALID."}
|
|
{"id": 13223, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "method", "why": "Merging method for RL-trained agentic models handling sparse task vectors"}], "pattern": "vary-conditions", "fixes": ["tasks"], "varies": ["merging method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Behavior knowledge merging tailored to sparse RL task vectors in agentic models."}
|
|
{"id": 9870, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "method", "why": "Optimal token baseline for variance reduction in long-horizon LLM RL"}], "pattern": "vary-conditions", "fixes": ["tasks"], "varies": ["baseline"], "benchmarks": [], "results": [], "workflow": ["Derive optimal baseline", "Logit-gradient proxy"], "summary": "OTB weights updates by cumulative gradient norm to stabilize long-horizon LLM RL."}
|
|
{"id": 14342, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "method", "why": "Frozen small LM probing for low-latency agent routing"}], "pattern": "vary-conditions", "fixes": ["agents"], "varies": ["router"], "benchmarks": [], "results": [{"benchmark": "enterprise agents", "system": "LatentGate", "base_model": null, "setting": "proposed", "metric": "in-domain accuracy", "score": 98.8, "unit": "%"}], "workflow": ["Extract hidden states", "PCA whitening", "Linear probe"], "summary": "LatentGate routes queries to specialized agents using frozen SLM hidden states."}
|
|
{"id": 3357, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Reverse-engineers website functionality into callable tools for web agents"}], "pattern": "vary-scaffold", "fixes": ["model"], "varies": ["tool interface"], "benchmarks": [], "results": [], "workflow": ["Discover site functions", "Create deterministic tools", "Agent calls tools"], "summary": "WALT converts website functionality into callable tools to simplify web agent tasks."}
|
|
{"id": 5730, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Neuro-symbolic LTL reward shaping with group automata for LLM agent RL"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["reward"], "benchmarks": [], "results": [], "workflow": ["Symbolize events", "Translate to LTL", "Shape rewards"], "summary": "GLARE shapes GRPO rewards via LTL automata over trajectory events."}
|