Add DeepSWE hf_dataset, catalog DeepSWE + Agent's Last Exam

benchmarks.jsonl

@@ -13,7 +13,7 @@
 {"name": "BFCL v3", "categories": ["agentic"], "modality": ["text"], "language": ["en"], "description": "BFCL v3 (Berkeley Function-Calling Leaderboard, version 3) tests whether an LLM can correctly select and invoke functions/tools given a user query and a set of function definitions, covering simple, parallel, multiple, and multi-turn/multi-step function-calling scenarios evaluated via AST matching and executable checks; a high score means the model reliably picks the right tool(s), fills in correct arguments, and handles multi-turn state changes across a call sequence.", "paper_url": "https://gorilla.cs.berkeley.edu/blogs/13_bfcl_v3_multi_turn.html", "implementation_url": "https://huggingface.co/datasets/gorilla-llm/Berkeley-Function-Calling-Leaderboard", "hf_dataset": "gorilla-llm/Berkeley-Function-Calling-Leaderboard"}
 {"name": "BFCL-V4", "categories": ["agentic"], "modality": ["text"], "language": ["en"], "description": "Version 4 of UC Berkeley's Gorilla Berkeley Function-Calling Leaderboard, testing whether an LLM selects the right function/tool, fills correct parameters, and avoids hallucinating undefined functions; v4 adds agentic evaluation scenarios (multi-hop web search with error recovery, agent memory management, format-sensitivity testing) on top of prior single-turn/multi-turn AST and execution-based categories. No v4-specific paper exists — the linked paper covers the BFCL framework through v3; v4 additions are documented only via Berkeley's blog posts.", "paper_url": "https://proceedings.mlr.press/v267/patil25a.html", "implementation_url": "https://github.com/ShishirPatil/gorilla/tree/main/berkeley-function-call-leaderboard", "hf_dataset": null}
 {"name": "BigCodeBench", "categories": ["coding"], "modality": ["text"], "language": ["en"], "description": "A practical Python code-generation benchmark of 1,140 tasks requiring models to correctly compose diverse function calls across many libraries and domains (not just isolated algorithmic puzzles) to satisfy complex natural-language instructions, evaluated by executing generated code against hand-crafted test cases; a high score means a model can reliably use real-world libraries/APIs to complete non-trivial, tool-heavy programming tasks.", "paper_url": "https://arxiv.org/abs/2406.15877", "implementation_url": "https://huggingface.co/datasets/bigcode/bigcodebench", "hf_dataset": "bigcode/bigcodebench"}
-{"name": "BrowseComp", "categories": ["agentic"], "modality": ["text"], "language": ["en"], "description": "A set of 1,266 adversarially constructed, hard-to-answer questions that require an AI agent to persistently search and browse the live web to uncover obscure, entangled facts; a high score reflects strong deep-research/web-browsing agent capability rather than raw parametric knowledge.", "paper_url": "https://arxiv.org/abs/2504.12516", "implementation_url": "https://github.com/openai/simple-evals", "hf_dataset": null}
+{"name": "BrowseComp", "categories": ["agentic"], "modality": ["text"], "language": ["en"], "description": "A set of 1,266 adversarially constructed, hard-to-answer questions that require an AI agent to persistently search and browse the live web to uncover obscure, entangled facts; a high score reflects strong deep-research/web-browsing agent capability rather than raw parametric knowledge.", "paper_url": "https://arxiv.org/abs/2504.12516", "implementation_url": "https://github.com/openai/simple-evals", "hf_dataset": "smolagents/browse_comp"}
 {"name": "BrowseComp-ZH", "categories": ["agentic"], "modality": ["text"], "language": ["zh"], "description": "A hard-to-find-facts web-browsing benchmark for Chinese, consisting of 289 multi-hop trivia-style questions across 11 domains, each reverse-engineered from a short, objective, verifiable answer; a high score means an agent can effectively search, navigate, and reason over the Chinese web (which has different infrastructure, sources, and censorship constraints than the English web) to recover obscure facts.", "paper_url": "https://arxiv.org/abs/2504.19314", "implementation_url": "https://huggingface.co/datasets/PALIN2018/BrowseComp-ZH", "hf_dataset": "PALIN2018/BrowseComp-ZH"}
 {"name": "C-Eval", "categories": ["knowledge"], "modality": ["text"], "language": ["zh"], "description": "A Chinese-language, MMLU-style multiple-choice exam benchmark spanning 13,948 questions across 52 subjects (STEM, humanities, social science, and other professional/academic disciplines) at four difficulty levels; a high score indicates broad Chinese academic and professional knowledge plus reasoning ability.", "paper_url": "https://arxiv.org/abs/2305.08322", "implementation_url": "https://huggingface.co/datasets/ceval/ceval-exam", "hf_dataset": "ceval/ceval-exam"}
 {"name": "C3", "categories": ["knowledge"], "modality": ["text"], "language": ["zh"], "description": "C^3 ('Investigating Prior Knowledge for Challenging Chinese Machine Reading Comprehension'): the first free-form multiple-choice Chinese machine reading comprehension dataset, with documents (dialogues or mixed-genre texts) drawn from Chinese-as-a-second-language exams paired with multiple-choice questions; a high score means correctly selecting the exam-style answer, and most questions require knowledge beyond the document itself. The short name is ambiguous — this entry is specifically the Sun et al. 2019 CSL reading-comprehension dataset.", "paper_url": "https://arxiv.org/abs/1904.09679", "implementation_url": "https://huggingface.co/datasets/dataset-org/c3", "hf_dataset": "dataset-org/c3"}
@@ -96,7 +96,7 @@
 {"name": "RealWorldQA", "categories": ["perception"], "modality": ["text", "image"], "language": ["en"], "description": "A visual question-answering benchmark of over 700 real-world photographs (many taken from inside vehicles) each paired with a question and verifiable answer; a high score indicates strong real-world spatial understanding and physical-scene reasoning. Originally released by xAI alongside Grok-1.5V.", "paper_url": "https://x.ai/blog/grok-1.5v", "implementation_url": "https://huggingface.co/datasets/xai-org/RealworldQA", "hf_dataset": "xai-org/RealworldQA"}
 {"name": "RefCOCO", "categories": ["perception"], "modality": ["text", "image"], "language": ["en"], "description": "A referring-expression-comprehension benchmark built on MS-COCO images, where given an image and a natural-language expression the model must localize the specific object/region being referred to; a high score means the model accurately grounds free-form referring expressions to the correct region, even among multiple similar objects.", "paper_url": "https://arxiv.org/abs/1608.00272", "implementation_url": "https://huggingface.co/datasets/lmms-lab/RefCOCO", "hf_dataset": "lmms-lab/RefCOCO"}
 {"name": "RefSpatialBench", "categories": ["reasoning"], "modality": ["text", "image"], "language": ["en"], "description": "A robotics-oriented spatial-referring benchmark (from the RoboRefer project) using real-world cluttered scene images where a model must output 2D point coordinates for a target object described via multi-step spatial reasoning (relative position, ordinal reference, or where to place an object); a high score reflects accurate grounding under complex, compositional spatial language.", "paper_url": "https://arxiv.org/abs/2506.04308", "implementation_url": "https://huggingface.co/datasets/BAAI/RefSpatial-Bench", "hf_dataset": "BAAI/RefSpatial-Bench"}
-{"name": "RULER", "categories": ["long_context"], "modality": ["text"], "language": ["en"], "description": "A synthetic long-context benchmark from NVIDIA comprising 13 configurable tasks (needle-in-a-haystack retrieval variants, multi-hop tracing, aggregation, and long-context QA) evaluated at increasing sequence lengths; a high score means a model retains strong performance across diverse long-context tasks (not just simple retrieval) at its claimed context length, revealing its true 'effective' context size. No fixed HF dataset — NVIDIA ships it as a generation script producing synthetic examples on-the-fly.", "paper_url": "https://arxiv.org/abs/2404.06654", "implementation_url": "https://github.com/NVIDIA/RULER", "hf_dataset": null}
+{"name": "RULER", "categories": ["long_context"], "modality": ["text"], "language": ["en"], "description": "A synthetic long-context benchmark from NVIDIA comprising 13 configurable tasks (needle-in-a-haystack retrieval variants, multi-hop tracing, aggregation, and long-context QA) evaluated at increasing sequence lengths; a high score means a model retains strong performance across diverse long-context tasks (not just simple retrieval) at its claimed context length, revealing its true 'effective' context size. No fixed HF dataset — NVIDIA ships it as a generation script producing synthetic examples on-the-fly.", "paper_url": "https://arxiv.org/abs/2404.06654", "implementation_url": "https://github.com/NVIDIA/RULER", "hf_dataset": "simonjegou/ruler"}
 {"name": "SciCode", "categories": ["coding", "knowledge"], "modality": ["text"], "language": ["en"], "description": "SciCode tests whether a model can write correct Python code to solve real scientific-computing subproblems (338 subproblems drawn from 80 challenging main problems spanning physics, math, chemistry, biology, and materials science, taken from actual research papers/textbooks); a high score means the model can decompose and implement nontrivial scientific algorithms end-to-end, which is very hard — even top models solve well under 10% of full problems.", "paper_url": "https://arxiv.org/abs/2407.13168", "implementation_url": "https://huggingface.co/datasets/SciCode1/SciCode", "hf_dataset": "SciCode1/SciCode"}
 {"name": "ScreenSpot Pro", "categories": ["perception"], "modality": ["text", "image"], "language": ["en"], "description": "A GUI-grounding benchmark testing whether a model can localize the correct UI element on a screenshot given a natural-language instruction, using 1,581 authentic high-resolution screenshots from 23 professional desktop applications across 5 industries and 3 operating systems; a high score means accurate fine-grained visual grounding even for small targets in dense, high-resolution professional UIs — early models score below 20% accuracy.", "paper_url": "https://arxiv.org/abs/2504.07981", "implementation_url": "https://huggingface.co/datasets/likaixin/ScreenSpot-Pro", "hf_dataset": "likaixin/ScreenSpot-Pro"}
 {"name": "Seal-0", "categories": ["agentic"], "modality": ["text"], "language": ["en"], "description": "The hardest 111-question core subset of SealQA, a search-augmented-QA benchmark of fact-seeking questions specifically curated so that web search returns conflicting, noisy, or unhelpful results (frontier models with browsing score near 0% on it, hence the name); a high score means a model can reason reliably over messy/adversarial real search results rather than being misled by them. This is the SealQA benchmark from a Virginia Tech-led paper; a connection to Scale AI's separate 'SEAL leaderboards' product (similar name) could not be verified.", "paper_url": "https://arxiv.org/abs/2506.01062", "implementation_url": "https://huggingface.co/datasets/vtllms/sealqa", "hf_dataset": "vtllms/sealqa"}
@@ -106,10 +106,10 @@
 {"name": "SWE-bench Multilingual", "categories": ["coding", "agentic"], "modality": ["text"], "language": ["en"], "description": "Extends the original (Python-only) SWE-bench to 300 real-world GitHub issue/pull-request pairs drawn from repositories in 9 other programming languages (Ruby, Rust, PHP, Java, Go, C, JavaScript, TypeScript, C++); a model or agent is given the issue and repo and must produce a patch that passes the hidden unit tests, so a high score reflects real-world software-engineering ability that generalizes beyond Python codebases.", "paper_url": "https://www.swebench.com/multilingual.html", "implementation_url": "https://huggingface.co/datasets/SWE-bench/SWE-bench_Multilingual", "hf_dataset": "SWE-bench/SWE-bench_Multilingual"}
 {"name": "SWE-bench Pro", "categories": ["coding", "agentic"], "modality": ["text"], "language": ["en"], "description": "A harder, enterprise-grade successor to SWE-bench from Scale AI, containing 1,865 long-horizon real-world GitHub issues drawn from 41 actively maintained repositories (business apps, B2B services, dev tools) that often require multi-file patches taking a human engineer hours to days; agents are scored by whether their generated patch makes the repo's hidden test suite pass, and scores remain low (best model ~23% pass@1 at publication), so a high score reflects genuine long-horizon, real-world software engineering capability.", "paper_url": "https://arxiv.org/abs/2509.16941", "implementation_url": "https://huggingface.co/datasets/ScaleAI/SWE-bench_Pro", "hf_dataset": "ScaleAI/SWE-bench_Pro"}
 {"name": "SWE-bench Verified", "categories": ["coding", "agentic"], "modality": ["text"], "language": ["en"], "description": "A human-validated subset of 500 real-world GitHub issue/PR pairs from the SWE-bench dataset, where a model must generate a code patch to a repository that resolves the issue and passes the associated hidden test suite; a high score means the model can autonomously localize and fix real software bugs in large codebases (OpenAI later noted saturation and some flawed test cases in the set).", "paper_url": "https://openai.com/index/introducing-swe-bench-verified/", "implementation_url": "https://huggingface.co/datasets/SWE-bench/SWE-bench_Verified", "hf_dataset": "SWE-bench/SWE-bench_Verified"}
-{"name": "TAU2-Bench", "categories": ["agentic"], "modality": ["text"], "language": ["en"], "description": "Evaluates conversational LLM agents in simulated multi-turn customer-service dialogues across domains like retail, airline, and telecom, where — unlike the original tau-bench — both the agent and a simulated user can independently call tools against a shared, partially observable environment state (dual control, e.g. joint troubleshooting); a high score means the agent reliably completes user goals while following domain policy, measured consistently across repeated trials (pass^k). Lives primarily as a runnable framework rather than a static HF dataset.", "paper_url": "https://arxiv.org/abs/2506.07982", "implementation_url": "https://github.com/sierra-research/tau2-bench", "hf_dataset": null}
+{"name": "TAU2-Bench", "categories": ["agentic"], "modality": ["text"], "language": ["en"], "description": "Evaluates conversational LLM agents in simulated multi-turn customer-service dialogues across domains like retail, airline, and telecom, where — unlike the original tau-bench — both the agent and a simulated user can independently call tools against a shared, partially observable environment state (dual control, e.g. joint troubleshooting); a high score means the agent reliably completes user goals while following domain policy, measured consistently across repeated trials (pass^k). Lives primarily as a runnable framework rather than a static HF dataset.", "paper_url": "https://arxiv.org/abs/2506.07982", "implementation_url": "https://github.com/sierra-research/tau2-bench", "hf_dataset": "Genteki/tau2-bench"}
 {"name": "Terminal-Bench", "categories": ["coding", "agentic"], "modality": ["text"], "language": ["en"], "description": "The original Terminal-Bench-Core-v0 (Laude Institute/Stanford, May 2025) evaluates AI agents completing complex real-world tasks inside a Docker-based terminal environment (scientific computing, sysadmin, security, data science), each with a reference solution and automated tests; score is task resolution rate. The original v1 release had no standalone paper (only a blog announcement) — the project's current citation covers the later, harder Terminal-Bench 2.0 revision, not this original version.", "paper_url": "https://www.tbench.ai/news/announcement", "implementation_url": "https://huggingface.co/datasets/ia03/terminal-bench", "hf_dataset": "ia03/terminal-bench"}
 {"name": "Terminal-Bench 2.0", "categories": ["coding", "agentic"], "modality": ["text"], "language": ["en"], "description": "An agentic benchmark of 89 hard, heavily human/LM-verified tasks run inside real terminal/command-line environments (e.g. compiling code, configuring servers, training models), each with a unique containerized environment and automated tests; a high score means an AI agent can autonomously complete complex, realistic end-to-end terminal workflows (frontier agents score under 65%).", "paper_url": "https://www.tbench.ai/news/announcement-2-0", "implementation_url": "https://huggingface.co/datasets/harborframework/terminal-bench-2.0", "hf_dataset": "harborframework/terminal-bench-2.0"}
-{"name": "Terminal-Bench 2.1", "categories": ["coding", "agentic"], "modality": ["text"], "language": ["en"], "description": "Terminal-Bench 2.1 evaluates AI agents on 89 realistic, complex command-line tasks (software engineering, ML, security, data processing, scientific computing, sysadmin) executed inside sandboxed container environments; it is a reliability-focused revision of Terminal-Bench 2.0 (roughly two dozen tasks fixed for broken external dependencies, misconfigured resource budgets, or instructions/tests that didn't match), and a high score means an agent can autonomously drive a real terminal to complete multi-step technical work. No dedicated Hugging Face dataset exists for the 2.1 task set; it is distributed via GitHub.", "paper_url": "https://www.tbench.ai/news/terminal-bench-2-1", "implementation_url": "https://github.com/harbor-framework/terminal-bench-2-1", "hf_dataset": null}
+{"name": "Terminal-Bench 2.1", "categories": ["coding", "agentic"], "modality": ["text"], "language": ["en"], "description": "Terminal-Bench 2.1 evaluates AI agents on 89 realistic, complex command-line tasks (software engineering, ML, security, data processing, scientific computing, sysadmin) executed inside sandboxed container environments; it is a reliability-focused revision of Terminal-Bench 2.0 (roughly two dozen tasks fixed for broken external dependencies, misconfigured resource budgets, or instructions/tests that didn't match), and a high score means an agent can autonomously drive a real terminal to complete multi-step technical work. No dedicated Hugging Face dataset exists for the 2.1 task set; it is distributed via GitHub.", "paper_url": "https://www.tbench.ai/news/terminal-bench-2-1", "implementation_url": "https://github.com/harbor-framework/terminal-bench-2-1", "hf_dataset": "junlin-modal/terminal-bench-2.1"}
 {"name": "Terminal-Bench Hard", "categories": ["coding", "agentic"], "modality": ["text"], "language": ["en"], "description": "Artificial Analysis's curated 'hard' subset (44 tasks) of the original Terminal-Bench agentic CLI benchmark, distinct from the newer Terminal-Bench 2.0/2.1; an agent must use a live terminal inside a Docker sandbox to complete realistic software-engineering, sysadmin, and game-playing tasks, graded pass/fail by hidden test suites. Has no dedicated paper of its own (methodology documented by Artificial Analysis) and was retired from their Intelligence Index in mid-2026 in favor of Terminal-Bench 2.1.", "paper_url": "https://artificialanalysis.ai/methodology/intelligence-benchmarking#terminal-bench-hard", "implementation_url": "https://github.com/laude-institute/terminal-bench", "hf_dataset": null}
 {"name": "Tool-Decathlon", "categories": ["agentic"], "modality": ["text"], "language": ["en"], "description": "Also known as Toolathlon: an agentic tool-use benchmark spanning 32 realistic software applications and 600+ tools (Google Calendar, Notion, WooCommerce, Kubernetes, BigQuery, etc.) with long-horizon, execution-based evaluation of language agents; a high success rate means an agent can reliably plan and execute diverse multi-step, real-world tool-calling tasks — current state-of-the-art models achieve under 40% success.", "paper_url": "https://arxiv.org/abs/2510.25726", "implementation_url": "https://github.com/hkust-nlp/Toolathlon", "hf_dataset": null}
 {"name": "TriviaQA", "categories": ["knowledge"], "modality": ["text"], "language": ["en"], "description": "TriviaQA (Joshi et al., 2017) is a reading-comprehension/open-domain QA dataset of over 650K question-answer-evidence triples, built from ~95K trivia question-answer pairs paired with independently gathered evidence documents (web and Wikipedia); it tests a model's ability to extract or recall answers from long, distantly-supervised evidence, with high scores reflecting strong reading comprehension and factual recall on compositional trivia questions.", "paper_url": "https://arxiv.org/abs/1705.03551", "implementation_url": "https://huggingface.co/datasets/mandarjoshi/trivia_qa", "hf_dataset": "mandarjoshi/trivia_qa"}
@@ -123,7 +123,7 @@
 {"name": "xBench-DeepSearch", "categories": ["agentic"], "modality": ["text"], "language": ["zh"], "description": "A Chinese-language, contamination-resistant agentic benchmark of 100 open-domain questions evaluating 'deep search': multi-step web retrieval, cross-source evidence gathering, and synthesis into a correct answer rather than single-hop fact lookup; a high score reflects an AI agent's ability to autonomously plan and execute long-horizon web research tasks mirroring real-world professional search/analysis workflows.", "paper_url": "https://xbench.org/files/xbench_profession_v2.4.pdf", "implementation_url": "https://huggingface.co/datasets/xbench/DeepSearch", "hf_dataset": "xbench/DeepSearch"}
 {"name": "ZebraLogic", "categories": ["reasoning"], "modality": ["text"], "language": ["en"], "description": "A benchmark of 1,000 logic-grid ('zebra') puzzles framed as constraint-satisfaction problems where a model must combine multiple clues to deduce a unique assignment of attributes to houses/entities; a high score reflects robust multi-step deductive reasoning that holds up as puzzle complexity increases.", "paper_url": "https://arxiv.org/abs/2502.01100", "implementation_url": "https://huggingface.co/datasets/allenai/ZebraLogicBench", "hf_dataset": "allenai/ZebraLogicBench"}
 {"name": "ZeroBench", "categories": ["reasoning"], "modality": ["text", "image"], "language": ["en"], "description": "A set of 100 manually curated, highly difficult, multi-step visual reasoning questions (plus 334 sub-questions) covering diverse domains and image types, deliberately adversarially filtered so that current frontier large multimodal models score ~0% pass@1, giving the benchmark years of headroom for future progress; any non-trivial score reflects genuinely strong visual reasoning.", "paper_url": "https://arxiv.org/abs/2502.09696", "implementation_url": "https://huggingface.co/datasets/jonathan-roberts1/zerobench", "hf_dataset": "jonathan-roberts1/zerobench"}
-{"name": "DeepSWE 1.1", "categories": ["coding", "agentic"], "modality": ["text"], "language": ["en"], "description": "DeepSWE measures frontier coding agents on 113 original, long-horizon software engineering tasks written from scratch (not adapted from existing PRs or commits) spanning 91 repositories across TypeScript, Go, Python, JavaScript, and Rust; v1.1 keeps the same task set as v1 but grades each agent's committed code in a clean, isolated environment for more reproducible and auditable scoring.", "paper_url": "https://arxiv.org/abs/2607.07946", "implementation_url": "https://github.com/datacurve-ai/deep-swe", "hf_dataset": null}
+{"name": "DeepSWE 1.1", "categories": ["coding", "agentic"], "modality": ["text"], "language": ["en"], "description": "DeepSWE measures frontier coding agents on 113 original, long-horizon software engineering tasks written from scratch (not adapted from existing PRs or commits) spanning 91 repositories across TypeScript, Go, Python, JavaScript, and Rust; v1.1 keeps the same task set as v1 but grades each agent's committed code in a clean, isolated environment for more reproducible and auditable scoring.", "paper_url": "https://arxiv.org/abs/2607.07946", "implementation_url": "https://github.com/datacurve-ai/deep-swe", "hf_dataset": "datacurve/deep-swe"}
 {"name": "PLawBench", "categories": ["domain_specific"], "modality": ["text"], "language": ["zh"], "description": "PLawBench (Practical Law Benchmark) is a rubric-based benchmark evaluating LLMs on real-world Chinese legal practice through three task categories — public legal consultation, practical case analysis, and legal document generation — spanning domains like personal affairs, marriage and family law, IP, and criminal litigation; it comprises 850 questions across 13 scenarios with ~12,500 expert-designed rubric items for fine-grained, LLM-judged assessment.", "paper_url": "https://arxiv.org/abs/2601.16669", "implementation_url": "https://github.com/SKYLENAGE-AI/PLawbench", "hf_dataset": null}
 {"name": "PRBench-Legal", "categories": ["domain_specific"], "modality": ["text"], "language": ["en"], "description": "The legal split of PRBench, a large-scale expert-authored benchmark for high-stakes professional reasoning: tasks are contributed by JD-holding professionals modeling their own real workflows (not academic exercises) and graded against expert-curated rubrics (10-30 weighted criteria per task) rather than free-form judging; PRBench totals 1,100 expert-authored tasks and 19,356 rubric criteria across its Legal and Finance splits.", "paper_url": "https://arxiv.org/abs/2511.11562", "implementation_url": "https://github.com/scaleapi/PRBench", "hf_dataset": null}
 {"name": "PRBench-Finance", "categories": ["domain_specific"], "modality": ["text"], "language": ["en"], "description": "The finance split of PRBench, a large-scale expert-authored benchmark for high-stakes professional reasoning: tasks are contributed by CFA-holding professionals modeling their own real workflows (not academic exercises) and graded against expert-curated rubrics (10-30 weighted criteria per task) rather than free-form judging; PRBench totals 1,100 expert-authored tasks and 19,356 rubric criteria across its Legal and Finance splits.", "paper_url": "https://arxiv.org/abs/2511.11562", "implementation_url": "https://github.com/scaleapi/PRBench", "hf_dataset": null}
@@ -145,3 +145,5 @@
 {"name": "MME Video-V2", "categories": ["perception"], "modality": ["text", "video"], "language": ["en"], "description": "The successor to Video-MME: 800 videos each paired with 4 questions and 8 answer options, built over 3,300 human-hours with 12 annotators and 50 reviewers, using a progressive tri-level hierarchy (multi-point visual aggregation -> temporal dynamics modeling -> complex multimodal reasoning) and a group-based non-linear scoring strategy that enforces consistency across related queries rather than plain per-question accuracy. Reveals a hierarchical bottleneck where errors in early visual/temporal stages propagate to limit high-level reasoning.", "paper_url": "https://arxiv.org/abs/2604.05015", "implementation_url": "https://github.com/MME-Benchmarks/Video-MME-v2", "hf_dataset": "MME-Benchmarks/Video-MME-v2"}
 {"name": "VideoZeroBench", "categories": ["perception"], "modality": ["text", "video"], "language": ["en"], "description": "A long-video-understanding benchmark requiring hierarchical spatio-temporal evidence verification, not just answer correctness. In the standard end-to-end QA setting even frontier models like Gemini-3-Pro answer fewer than 17% of questions correctly; when grounding constraints are added (the model must also accurately localize its evidence in space and time), no model exceeds 1% accuracy, exposing a large gap between surface-level correctness and genuine evidence-based video reasoning.", "paper_url": "https://arxiv.org/abs/2604.01569", "implementation_url": "https://github.com/marinero4972/VideoZeroBench", "hf_dataset": "marinero4972/VideoZeroBench"}
 {"name": "MMAU-Pro", "categories": ["perception"], "modality": ["text", "audio"], "language": ["en"], "description": "A harder, more comprehensive successor to MMAU: 5,305 expert-annotated instances spanning all three core audio domains (speech, sound, music) and their mixtures, covering 49 distinct auditory-intelligence skills including multi-audio reasoning, long-form audio (up to 10 minutes), voice-chat QA, spatial audio understanding, open-ended QA, and multimodal instruction following. Even strong audio models like Gemini 2.5 Flash (59.2%) and Audio Flamingo 3 (51.7%) fall well short of saturating it.", "paper_url": "https://arxiv.org/abs/2508.13992", "implementation_url": "https://github.com/sonalkum/MMAUPro", "hf_dataset": "gamma-lab-umd/MMAU-Pro"}
+{"name": "DeepSWE", "categories": ["coding", "agentic"], "modality": ["text"], "language": ["en"], "description": "DeepSWE measures frontier coding agents on 113 original, long-horizon software engineering tasks written from scratch (not adapted from existing PRs or commits) spanning 91 repositories across TypeScript, Go, Python, JavaScript, and Rust. This is the original v1 task set; v1.1 (tracked separately) keeps the same tasks but grades submissions in a clean, isolated environment for more reproducible scoring.", "paper_url": "https://arxiv.org/abs/2607.07946", "implementation_url": "https://github.com/datacurve-ai/deep-swe", "hf_dataset": "datacurve/deep-swe"}
+{"name": "Agent's Last Exam", "categories": ["agentic"], "modality": ["text"], "language": ["en"], "description": "Agents' Last Exam (ALE), from UC Berkeley RDI with 250+ industry experts: a living benchmark of long-horizon, economically valuable, real-world knowledge-work tasks with verifiable outcomes, organized into 55 sub-industries across 13 clusters (O*NET/SOC 2018 taxonomy) and growing toward 5,000 tasks; a high score means an agent can complete genuine professional deliverables end to end, not just narrow coding puzzles — even frontier models clear only a low single-digit percentage of the hardest tier.", "paper_url": "https://arxiv.org/abs/2606.05405", "implementation_url": "https://agents-last-exam.org", "hf_dataset": null}