Benchmark2606.09426
WeaveBench: A Long-Horizon, Real-World Benchmark for Computer-Use Agents with Hybrid Interfaces
114 tasks requiring GUI+CLI+code orchestration; best PassRate 41.2%.
This day in AI
Recent paper calendar for people tracking useful AI shifts.
CodeSOTA scans arXiv days for papers that could become useful benchmark rows, model-selection evidence, or product-facing research notes. This page keeps the daily trail visible instead of burying it in one-off reports.
Calendar
Research days worth revisiting
Tuesday
June 9, 2026
Today's batch is dominated by new benchmarks for agentic systems, safety, and multimodal reasoning, alongside a strong signal for structured evaluation and self-evolving architectures.
509 entries
Monday
June 8, 2026
Agentic self-evolution, formal verification, safety evaluations, and new domain-specific benchmarks dominate Monday's batch, revealing growing rigor in measuring and improving autonomous systems.
164 entries
Friday
June 5, 2026
A dense crop of task-specific benchmarks and agent systems reveals that frontier models saturate narrow perception tests but still fail on long-horizon, stateful, and multi-modal safety reasoning, while new training and inference techniques show materials for closing those gaps.
280 entries
Thursday
June 4, 2026
Benchmarks and systems for agent safety, long-horizon code optimization, consequence-aware compute allocation, and structured skill learning dominate today's cs.AI submissions.
207 entries
Wednesday
June 3, 2026
Today's batch is dominated by rigorous new benchmarks that expose agent fragility in finance, medicine, and software engineering, alongside system papers that improve evaluation efficiency and safety.
440 entries, including 80 new submissions
Wednesday
May 20, 2026
New benchmarks for agentic routing, long-horizon software development, memory evaluation, and healthcare workflows reveal persistent gaps in frontier models, while dynamic layer routing offers a path to more efficient LLM inference.
460 entries, including 70 new submissions
Tuesday
May 19, 2026
The day was dominated by agent runtime security, process-aware benchmarks, self-improving agent systems, and sparse reasoning credit assignment.
900 entries, including 142 new submissions
Monday
May 18, 2026
Monday's useful signal was practical agent infrastructure: SaaS and shopping environments, formal monitoring, research-agent scaffolding, and vertical medical systems.
200 entries, sampled LLM scout plus full deterministic screen
Friday
May 15, 2026
Friday's signal was about making agent behavior legible: orchestration graphs, memory cold starts, design patterns, education agents, and benchmarked governance workflows.
445 entries, including 100 new submissions
Today's batch is dominated by new benchmarks for agentic systems, safety, and multimodal reasoning, alongside a strong signal for structured evaluation and self-evolving architectures.
Big picture
- Agent benchmarks expand to long-horizon, hybrid-interface, and adversarial settings
- Safety and alignment work targets emergent misalignment, compliance, and overconfidence
- Structured evaluation frameworks for tabular, medical, and formal reasoning domains
- Self-evolution and influence-guided training for reasoning improvement
Benchmarks to extract
- WeaveBench: extract PassRate and trajectory-aware judge correlation on hybrid tasks
- KernelBench: confirm attack success rate reduction from 62% to 0% after hacker-fixer loop
- TRL-Bench: verify that generic text encoders lead on surface-text tasks, tabular specialists on aligned tasks
- IMUG-Bench: confirm exposure bias gap between understanding and generation in multi-turn settings
Papers and links
Safety2606.08960
Hardening Agent Benchmarks with Adversarial Hacker-Fixer Loops
Hacker-fixer loop drives attack success from 62% to 0% on KernelBench.
Reasoning2606.09052
INFUSER: Influence-Guided Self-Evolution Improves Reasoning
Co-training framework with DuGRPO; 20% relative gain on Olympiad/SuperGPQA.
Benchmark2606.09323
TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular Encoders
20 models, 16 tasks; shows encoder quality is capability-specific.
Benchmark2606.09169
IMUG-Bench: Benchmarking Unified Multimodal Models on Interleaved Understanding and Generation
3,113 samples across static, temporal, and hybrid tasks; exposes exposure bias.
System2606.08106
PACE: Anytime-Valid Acceptance Tests for Self-Evolving Agents
Training-free commit gate reduces false commits from 30-42% to near zero.
Method note
Sampled 80 deterministic candidates from 509 cs.AI entries, prioritized by artifact names, benchmark language, and quantitative evidence, then selected 6 papers covering agent benchmarks, safety, reasoning, and structured evaluation. Abstract-level summaries are triage, not final benchmark claims; papers with tables or model rankings still need PDF/source extraction before they become CodeSOTA evidence rows.
Agentic self-evolution, formal verification, safety evaluations, and new domain-specific benchmarks dominate Monday's batch, revealing growing rigor in measuring and improving autonomous systems.
Big picture
- Closed-loop self-evolution frameworks that reuse agent traces for targeted task generation
- Formal modeling and verification of agent workflows using dependent-type languages
- Strategic attack selection as a critical blind spot in agentic AI control evaluations
- A surge of domain-specific benchmarks covering macOS GUIs, drug design, and mathematical research discussions
Benchmarks to extract
- Socratic-SWE: exact SWE-bench Verified score per iteration (abstract reports 50.40% after 3 rounds)
- MacArena: per-model accuracy on macOS-native subset (abstract reports leader trailing >26%)
- Attack Selection: exact safety reduction in pp at 1% audit (abstract reports 20pp on BashArena and LinuxArena start policy)
- Lean4Agent: verification-passing vs. failing workflow accuracy on SWE-Bench-Verified subset (abstract reports 11.94% average difference)
Papers and links
Agent2606.07412
Socratic-SWE: Self-Evolving Coding Agents via Trace-Derived Agent Skills
Closed-loop self-evolution distills solving traces into structured skills to generate targeted repair tasks; achieves 50.40% on SWE-bench Verified after three iterations.
Benchmark2606.06560
MacArena: Benchmarking Computer Use Agents on an Online macOS Environment
421 manually verified tasks (50 apps) on native Apple Silicon; model rankings invert between ported and macOS-native tasks (leader trails by >26%), revealing cross-platform GUI competence gaps.
Safety2606.06529
Attack Selection in Agentic AI Control Evaluations Meaningfully Decreases Safety
Strategic start/stop attack policies reduce measured safety by 20–28pp at 1% audit budget on BashArena and LinuxArena, suggesting control evaluations may be overly optimistic.
System2606.06523
Lean4Agent: Formal Modeling and Verification for Agent Workflow and Trajectory
First Lean4-based framework for agent behavior verification; verified workflows outperform failing ones by 11.94% on SWE-Bench-Verified and ELAIP-Bench, with 7.47% further gain from LeanEvolve revision.
Benchmark2606.07462
Act As a Real Researcher: A Suite of Benchmarks Evaluating Frontier LLMs and Agentic Harnesses in Research Lifecycle
AARRI-Bench targets granular research behavior; best system (Mini-SWE-Agent + Claude Opus 4.7) achieves only 68.3%, frequently missing subtle details humans catch.
Benchmark2606.06717
ShallowBench: Benchmarking Generative Drug Design Models on Shallow-Pocket Targets
5,780 shallow-pocket targets from CrossDocked2020; SOTA generative models show weaker predicted binding affinity on low-concavity interfaces (e.g., KRAS, MYC), highlighting a key failure mode.
Method note
Collection of the 164-entry cs.AI June 8, 2026 recent-day listing via arXiv API, deterministic sampling of the top 77 benchmark/SOTA-signal candidates, followed by manual selection of six papers balancing agent systems, safety, and domain-specific benchmarks. Abstract-level summaries are triage, not final benchmark claims; papers with tables or model rankings still need PDF/source extraction before they become CodeSOTA evidence rows.
A dense crop of task-specific benchmarks and agent systems reveals that frontier models saturate narrow perception tests but still fail on long-horizon, stateful, and multi-modal safety reasoning, while new training and inference techniques show materials for closing those gaps.
Big picture
- Long-running / stateful agent evaluation emerges as a core challenge (SentinelBench, CL-Bench, Agents' Last Exam)
- Code-agent benchmarks shift from completion to repository-level tool-use and framework evaluation (TensorBench, ADK Arena)
- Safety and jailbreak detection advance with both system-level guards and multi-modal safety benchmarks (GuardNet, MCBench, SlotGCG)
- New training recipes (GRPO variants, CKA distillation, hypernetwork adapters) promise more efficient and reliable reasoning and control
Benchmarks to extract
- SentinelBench task completion and reaction-time tradeoffs (table in §4)
- CL-Bench gain metric isolating online learning from model capability
- TensorBench pass rates and pairwise Cohen's κ across 7 agents
- MCBench cross-modal safety accuracy breakdown by modality combination and risk category
Papers and links
Benchmark2606.05342
SentinelBench: A Benchmark for Long-Running Monitoring Agents
100 tasks across 10 synthetic web environments test whether agents can monitor, wait, and react promptly over time rather than acting continuously.
Benchmark2606.05661
Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments
First expert-validated benchmark for continual learning across 6 domains; naive ICL outperforms dedicated memory systems, exposing headroom for better stateful architectures.
Benchmark2606.05570
TensorBench: Benchmarking Coding Agents on a Compiler-Based Tensor Framework
199 repo-level tasks on a real PyTorch extension; strongest agent passes 64.8%, and pairwise agreement between agents is low (κ=0.05), revealing task-specific skill gaps.
Safety2606.05177
MCBench: A Multicontext Safety Assessment Benchmark for Omni Large Language Models
1196 scenarios across vision, audio, and text; current Omni LLMs fail to integrate cross-modal cues for safety judgments, performing better only when salient signals are present.
Agent2606.05922
Retrospective Harness Optimization: Improving LLM Agents via Self-Preference over Trajectory Rollouts
Self-supervised method that re-solves past tasks and selects harness updates by pairwise self-preference; improves SWE-Bench Pro pass rate from 59% to 78% without external labels.
Reasoning2606.06468
Goedel-Architect: Streamlining Formal Theorem Proving with Blueprint Generation and Refinement
Uses blueprint dependency graphs and parallel Lean proving with DeepSeek-V4-Flash; achieves 99.2% pass@1 on MiniF2F-test and 75.6% on PutnamBench at a fraction of prior cost.
Method note
Collected all 280 cs.AI entries from the June 5 arXiv recent-day listing, filtered 80 deterministic candidates with strong benchmark or SOTA language, sampled metadata for 60, and selected 6 papers with the highest combined artifact signal, domain diversity, and practical impact on agent/benchmark/safety evaluation. Abstract-level summaries are triage, not final benchmark claims; papers with tables or model rankings still need PDF/source extraction before they become CodeSOTA evidence rows.
Benchmarks and systems for agent safety, long-horizon code optimization, consequence-aware compute allocation, and structured skill learning dominate today's cs.AI submissions.
Big picture
- Agent safety and intervention timing: new benchmarks for AI companion safety and a critical look at LLM-judge-based intervention triggers
- Long-horizon and consequence-aware reasoning: benchmarks for iterative code optimization and compute allocation based on task cost
- Structured skill and memory for agents: graph-based skill protocols, parametric memory for self-evolving agents, and state-grounded skill retrieval
- Cross-domain and multi-modal benchmarks: unified CAD, end-to-end autonomous driving, and cybersecurity vulnerability lifecycle evaluation
Benchmarks to extract
- AICompanionBench: per-category accuracy of LLM-as-judge on 9 safety risk categories (e.g., manipulation, self-harm)
- AutoLab: task completion rate under wall-clock budget across 36 long-horizon optimization tasks
- CyberGym-E2E: end-to-end success rate on vulnerability discovery, PoC generation, and patch generation for 920 CVEs
- SWE-bench Lite: cost-weighted loss reduction under consequence-aware vs. difficulty-aware compute allocation
Papers and links
Benchmark2606.04867
AICompanionBench: Benchmarking LLMs-as-Judges for AI Companion Safety
First public benchmark (2,123 real Replika conversations) for fine-grained safety risk categories; reveals LLMs struggle with implicit unsafe interactions like manipulation.
System2606.04402
Not All Errors Are Equal: Consequence-Aware Reasoning Compute Allocation
Proposes consequence-aware test-time compute allocation; on SWE-bench Lite, reduces cost-weighted loss by 22–33% vs. difficulty-only routing.
Benchmark2606.05080
AutoLab: Can Frontier Models Solve Long-Horizon Auto Research and Engineering Tasks?
36 expert-curated tasks for ultra long-horizon closed-loop optimization; finds persistence—not initial quality—is the dominant success predictor.
Agent2606.04536
Scaling Self-Evolving Agents via Parametric Memory
Introduces TMEM: agents absorb distilled supervision into fast LoRA weights via online updates, outperforming summary/retrieval baselines on LoCoMo and LongMemEval.
System2606.04781
AIP: A Graph Representation for Learning and Governing Agent Skills
Models skills as directed execution graphs; compiling human-written skills to AIP raised Claude Sonnet's mean task reward from 0.60 to 0.71 on SkillsBench.
Benchmark2606.04460
CyberGym-E2E: Scalable Real-World Benchmark for AI Agents' End-to-End Cybersecurity Capabilities
920 real-world vulnerabilities across 139 open-source projects; evaluates full lifecycle of discovery, PoC generation, and patch generation.
Method note
Sampled 60 of 207 cs.AI entries from the June 4 arXiv listing, then selected 6 papers with the strongest benchmark, agent, safety, or infrastructure signals based on artifact names, known benchmarks, and SOTA language in titles and abstracts. Abstract-level summaries are triage, not final benchmark claims; papers with tables or model rankings still need PDF/source extraction before they become CodeSOTA evidence rows.
Today's batch is dominated by rigorous new benchmarks that expose agent fragility in finance, medicine, and software engineering, alongside system papers that improve evaluation efficiency and safety.
Big picture
- Agent benchmarks for high-stakes domains: finance (BigFinanceBench, Hedge-Bench) and clinical software (MedCUA-Bench) show frontier models score below 60%.
- Process-level evaluation gains traction: AgentLens reveals 10.7% of passing SWE-agent trajectories are 'Lucky Passes' that tests miss.
- Resource-efficient evaluation: ProEval uses transfer learning to estimate performance with 8-65x fewer samples; TriEval runs bias/toxicity checks on a laptop.
- Safety and abstention: AgentRedBench cuts indirect prompt injection ASR from 69.9% to 2.4%; a new taxonomy formalizes when agents should refuse to act.
Benchmarks to extract
- Extract BigFinanceBench rubric scores per workflow step to localize where agents fail (best system 58.8%).
- Confirm Hedge-Bench deterministic grading methodology and whether the <16% score holds for GPT-5 and Claude Opus 4.
- Verify AgentLens-Bench 'Lucky Pass' rate (10.7%) and whether model rankings shift by 5+ positions when using quality score vs. pass rate.
- Extract AgentRedBench ASR per model and per attack type to confirm the 2.4% ASR with AgentRedGuard at 0.37% FPR.
Papers and links
Benchmark2606.03829
BigFinanceBench: A Workflow-Grounded Benchmark for Financial-Research Agents
928 expert-authored financial tasks with point-weighted rubrics; best agent scores only 58.8%.
Benchmark2606.03918
Hedge-Bench: Benchmarking Agents on Hard, Realistic Tasks Pertaining to Financial Reasoning
102 real hedge-fund analyst tasks with deterministic grading; frontier models score below 16%.
Benchmark2606.03203
MedCUA-Bench: A Screenshot-Only Benchmark for Clinical Computer-Use Agents
18 clinical scenarios across 10 domains; best closed-source model reaches 54.2% strict success, open-source agents average 2.5%.
Benchmark2605.12925
AgentLens: Revealing The Lucky Pass Problem in SWE-Agent Evaluation
Process-level analysis of 2,614 SWE-agent trajectories; 10.7% of passing runs are 'Lucky Passes' with chaotic behavior.
System2604.23099
ProEval: Proactive Failure Discovery and Efficient Performance Estimation for Generative AI Evaluation
Transfer-learning framework using Gaussian Processes; estimates performance within 1% of ground truth with 8-65x fewer samples.
Safety2606.02240
AgentRedBench: Dynamic Redteaming and Integration-Aware Defense for LLM Agents over SaaS Integrations
215 indirect prompt injection scenarios across 24 enterprise integrations; guard model cuts attack success rate from 69.9% to 2.4%.
Method note
Sampled 60 high-signal entries from 440 total (80 new, 203 cross-list, 157 replacements) by prioritizing papers with named benchmarks, SOTA claims, or safety/agent artifacts in title or abstract, then selected 6 for the digest. Abstract-level summaries are triage, not final benchmark claims; papers with tables or model rankings still need PDF/source extraction before they become CodeSOTA evidence rows.
New benchmarks for agentic routing, long-horizon software development, memory evaluation, and healthcare workflows reveal persistent gaps in frontier models, while dynamic layer routing offers a path to more efficient LLM inference.
Big picture
- Agentic routing and delegation benchmarks surface fundamental limitations in current orchestration methods
- Long-horizon software development and memory evaluation tasks expose sharp drops in performance as context grows
- Domain-specific benchmarks for healthcare and engineering construction highlight the need for specialized evaluation
- System-level efficiency gains from dynamic layer routing demonstrate progress in adaptive inference
Benchmarks to extract
- Verify that TwinRouterBench highest success rate is 64.8% for computer-use models.
- Verify that DecisionBench routing fidelity-at-1 ranges from 7.5% to 29.5% across conditions.
- Verify that RoadmapBench Claude-Opus-4.7 resolves only 39.1% of tasks.
- Verify that MINTEval average accuracy across all systems is 27.9%.
Papers and links
Benchmark2605.18859
TwinRouterBench: Fast Static and Live Dynamic Evaluation for Realistic Agentic LLM Routing
Step-level LLM routing benchmark with static and dynamic tracks for agentic workflows
Benchmark2605.19099
DecisionBench: A Benchmark for Emergent Delegation in Long-Horizon Agentic Workflows
Emergent delegation evaluation across GAIA, BFCL, and tau-bench
Benchmark2605.15846
RoadmapBench: Evaluating Long-Horizon Agentic Software Development Across Version Upgrades
115 long-horizon coding tasks from real version upgrades across 17 repos
Benchmark2605.18565
MINTEval: Evaluating Memory under Multi-Target Interference in Long-Horizon Agent Systems
15.6k QA pairs over long contexts averaging 138.8k tokens for multi-target memory
Benchmark2605.16679
CHI-Bench: Can AI Agents Automate End-to-End, Long-Horizon, Policy-Rich Healthcare Workflows?
End-to-end healthcare workflow automation with 20 apps and 87 MCP tools
System2510.12773
Dr.LLM: Dynamic Layer Routing in LLMs
Dynamic layer routing with MCTS-supervised per-layer routers for efficient LLM inference
Method note
Sampled 60 of 460 entries, prioritizing benchmarks and systems with the strongest quantitative signal from a deterministic candidate ranking. Abstract-level summaries are triage, not final benchmark claims; papers with tables or model rankings still need PDF/source extraction before they become CodeSOTA evidence rows.
The day was dominated by agent runtime security, process-aware benchmarks, self-improving agent systems, and sparse reasoning credit assignment.
Big picture
- Runtime agent safety is moving from prompt policies to action interception, MCP monitoring, and host-side controls.
- Evaluation is shifting toward process-aware tasks: tool trajectories, delivered artifacts, multimodal verification, and human-validated rubrics.
- Self-improving agents are becoming governed systems with rollback, canary tests, experience graphs, and explicit lifecycle controls.
- Reasoning work is converging on sparse credit assignment: find the decision tokens or reasoning steps that actually steer the answer.
Benchmarks to extract
- TOBench for tool-using agent rows
- ADR-Bench and SLEIGHT-Bench for agent security
- WebGameBench for coding-agent delivery
- LinAlg-Bench, CAM-Bench, and GIM for reasoning diagnostics
Papers and links
Benchmark2605.16909
TOBench: A Task-Oriented Omni-Modal Benchmark for Real-World Tool-Using Agents
100 executable tasks, 27 MCP servers, 324 tools, and closed-loop multimodal verification for end-to-end tool use.
Safety2605.17380
ADR: An Agentic Detection and Response System
Production-style monitoring for MCP agent activity with ADR-Bench covering 302 tasks and 17 attack techniques.
Benchmark2605.17637
WebGameBench: Requirement-to-Application Evaluation for Coding Agents
Evaluates browser-accessible delivered games, separating minimum working delivery from excellent requirement satisfaction.
Benchmark2605.16675
LinAlg-Bench: A Forensic Benchmark Revealing Structural Failure Modes
660 SymPy-certified linear algebra problems plus a failure taxonomy for diagnosing mathematical reasoning.
Reasoning2605.16874
Reasoning Can Be Restored by Correcting a Few Decision Tokens
Claims reasoning failures concentrate in a small number of early tokens, useful for intervention and evaluation design.
Agent2605.17721
EXG: Self-Evolving Agents with Experience Graphs
Turns successes and failures into graph memory, giving self-evolving agents a more inspectable substrate.
Method note
Full arXiv /new batch collected on May 19. The LLM scout covered all new submissions; deterministic benchmark detection covered the full batch. Abstract-level summaries are triage, not final benchmark claims; papers with tables or model rankings still need PDF/source extraction before they become CodeSOTA evidence rows.
Monday's useful signal was practical agent infrastructure: SaaS and shopping environments, formal monitoring, research-agent scaffolding, and vertical medical systems.
Big picture
- Real-world web agents are getting eval environments that look more like actual SaaS and commerce workflows.
- Agent architecture work is becoming cost-aware: context, hierarchy, reasoning depth, and monitoring are treated as budgeted design choices.
- Formal methods are appearing as runtime guardrails for LLM systems rather than only offline verification work.
- Medical and robotics papers are packaging open systems around concrete downstream workflows instead of generic model releases.
Benchmarks to extract
- ShopGym and SaaS-Bench for web-agent task pages
- PAGER/PAGE Bench for long-form or page-level agent evaluation
- ToxiAlert-Bench and RoadmapBench from the deterministic screen
- VLA-AD and RTL-BenchMT for embodied and hardware-facing rows
Papers and links
Benchmark2605.15777
SaaS-Bench
A practical benchmark direction for agents operating across SaaS workflows, useful for procurement-style agent evaluation.
Benchmark2605.16116
ShopGym
High-priority e-commerce agent environment surfaced by both the scout and deterministic screen.
Agent2605.16233
FORGE
Agent architecture signal around designing and evaluating practical multi-step systems.
Safety2605.16198
Formal Methods Meet LLMs
Runtime monitoring and formal constraints as a concrete control layer for LLM agents.
Agent2605.16217
Argus Deep Research Agents
Deep-research agent work that matches CodeSOTA's interest in paper-to-evidence workflows.
System2605.16215
Fully Open Meditron
Open medical AI system surfaced as a practical extraction target for model, data, and benchmark claims.
Method note
The dated May 18 arXiv /recent section was collected in full. We stopped the full LLM run after your subsample instruction and used a 60-paper scout plus deterministic screening across all 200 summaries. Abstract-level summaries are triage, not final benchmark claims; papers with tables or model rankings still need PDF/source extraction before they become CodeSOTA evidence rows.
Friday's signal was about making agent behavior legible: orchestration graphs, memory cold starts, design patterns, education agents, and benchmarked governance workflows.
Big picture
- Agent orchestration papers are moving toward explicit graphs, patterns, and inspectable memory instead of one-off prompt chains.
- Education and industrial benchmarks are becoming stronger examples of domain-specific agent evaluation.
- Safety and governance papers are trying to separate the task being performed from the governance process wrapped around it.
- Reasoning papers continue to probe symbolic structure, attributes, and limits of model-based inference.
Benchmarks to extract
- EntityBench, ClawForge, EduAgentBench, and Herculean
- PDI-Bench, Collider-Bench, and XDomainBench
- EduFrameTrap for sycophancy and education-agent failure modes
- SimPersona for persona or simulation-agent evaluation
Papers and links
Agent2605.13848
GraphBit
Agent orchestration and reliability signal around graph-structured coordination.
Agent2605.13880
PREPING
Agent memory cold-start work; useful for understanding reusable memory in deployed agents.
Benchmark2605.14133
ClawForge
High-priority benchmark candidate from the Friday benchmark report.
Benchmark2605.14322
EduAgentBench
Education-agent evaluation target with obvious CodeSOTA task-page value.
Benchmark2605.15199
EntityBench
Entity-centered benchmark surfaced as a strong extraction candidate.
Safety2605.14744
Governance-task decoupling
Useful framing for separating operational task success from oversight and governance quality.
Method note
This uses the existing May 15 batch and reports already in the local paper pipeline. Claims here are abstract/report level until tables are extracted from individual PDFs. Abstract-level summaries are triage, not final benchmark claims; papers with tables or model rankings still need PDF/source extraction before they become CodeSOTA evidence rows.
Next extraction pass
Turn the calendar into benchmark evidence, not just reading notes.
The valuable follow-up is to pull benchmark tables, model lists, task definitions, and failure taxonomies from the strongest papers. That gives CodeSOTA rows users can compare, not merely links they can browse.