Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity

• Seed2.0 extends the Seed model family and supports an ecosystem serving hundreds of millions of daily active users across applications.The broader family includes general-purpose, multimodal, open-source, code-specialized, diffusion, theorem-proving, and generative-media systems.Source: §1 · p. 1
• Seed2.0 targets production user experience through robust visual understanding, flexible inference across Pro, Lite, and Mini sizes, reliable complex instruction execution, and seamless coding assistance.These priorities reflect visual-query prevalence, latency effects, multi-step constraint satisfaction, and coding demand in online deployment.Source: §1 · p. 1
• Seed2.0 addresses real-world complexity by pursuing research-level reasoning, including Erdős problems and Scientific Coding, while systematically ingesting long-tail domain knowledge.The motivation is that current agents can solve competition-level problems yet often fail to complete practical end-to-end tasks because workflows are long-horizon and knowledge is domain-specific.Source: §1 · p. 1
• The report establishes a four-dimensional evaluation framework—Science Discovery, Vibe Coding, Context Learning, and Real-World Tasks—to track and guide complex, long-horizon agent performance.The framework serves both as a benchmark suite and as an iterative development guide.Source: §1 · p. 2
• Seed2.0 still has considerable gaps with Claude in coding and relatively obvious gaps with Gemini in long-tail knowledge, using SWE-Evo, NL2Repo, SuperGPQA, and SimpleQA-Verified as examples.The report presents these gaps as directions for improving real-world complexity capabilities.Source: §1 · p. 2

2.1 MaaS Usage in Mainland China

• Enterprise MaaS usage concentrates in digital industries and cognitively intensive applications, with Internet-sector traffic dominating and unstructured information processing representing the largest scenario.These deployments require heterogeneous-data processing, cross-domain synthesis, multi-step instructions, long contexts, structured generation, and tool-augmented execution.Source: §2.1 · pp. 2–3

2.2 Query Distribution in Agentic Coding

• Frontend development dominates agentic coding requests, with frontend languages comprising most touched code and Vue.js adoption exceeding React by more than three times.Bug fixing is the leading task type, followed by refactoring and documentation, indicating strong demand for reactive maintenance and debugging assistance.Source: §2.2 · p. 3, Figure 2 · p. 3

2.3 Cost Efficiency

• Seed2.0’s token pricing is roughly an order of magnitude lower than frontier models while delivering comparable user experience, enabling economically viable enterprise MaaS deployments.This advantage is especially relevant for high-volume, workflow-integrated tasks such as unstructured information processing and content generation.Source: Table 1 · p. 3, §2.3 · p. 4
• Seed2.0 offers Pro, Lite, and Mini tiers for complex reasoning, general-purpose applications, and high-throughput or latency-sensitive workloads, respectively.Mini’s decode pricing is under $0.50 per million tokens.Source: §2.3 · p. 4

3.1 Fundamental Language Capacity

• The framework evaluates Seed2.0’s language capabilities across reasoning, complex instruction following, broad knowledge, coding, science, and internal real-world benchmarks against frontier models.Evaluations include AIME 2025, HMMT 2025, BeyondAIME, IMOAnswerBench, AetherCode, LiveCodeBench, Codeforces, GPQA Diamond, PhyBench, ARC-AGI, MMLU-Pro, SuperGPQA, LPFQA, and others.Source: §3.1 · p. 4

3.1.1 Long-tail Professional Knowledge Benchmarks

• LPFQA, Encyclo-K, and HLE-Verified target reliable professional long-tail knowledge through realistic workplace questions, book-level structured knowledge, and expert-verified challenging problems.LPFQA covers programming, finance, engineering, medicine, and applied science; Encyclo-K supports zero-shot and few-shot evaluation; HLE-Verified filters unclear or unverifiable questions.Source: §3.1.1 · pp. 4–5

3.2 Fundamental Vision Capacity

• Seed2.0’s vision evaluation spans 50 image and 24 video benchmarks across mathematical, STEM, perceptual, spatial, document, long-context, reasoning, motion, and streaming capabilities.The framework specifies category-specific metrics and settings, including worst-case accuracy for DynaMath, normalized scoring for VibeEval, and adjusted frame rates for video benchmarks.Source: §3.2 · pp. 5–6

3.3 Fundamental Agentic Capacity

• Fundamental Agentic Capacity measures planning, tool use, environmental interaction, and multi-step completion across coding, search, tool invocation, GUI, and deep-research tasks.The evaluation adapts benchmark tasks to an internal agent framework, generally without external tools, while preserving original criteria and excluding three Terminal-Bench 2.0 cases for network and security constraints.Source: §3.3 · pp. 6–7

3.4 Advanced Economically & Scientifically Valuable Tasks

• Advanced evaluations cover Scientific Discovery, Vibe Coding, Context Learning, and Real-World Tasks through scenario-grounded benchmarks for research, repository construction, noisy-document reasoning, and executable planning.They include Ainstain Bench, BABE, NL2Repo-Bench, DeR2, GDPVal-Verified, XpertBench, and WorldTravel [107].Source: §3.4 · pp. 7–8
• The advanced suite also evaluates economically valuable workflows including education, text classification, information extraction, customer-support Q&A, and complex workflow completion.Overall, the suite measures whether models complete realistic, long-horizon tasks rather than merely answer isolated questions.Source: §3.4 · pp. 7–8

4.1 Fundamental Language Evaluation

• Seed2.0 Pro achieves top-tier language performance, leading on health knowledge, matching or surpassing frontier models on science, and excelling in mathematics, code, and STEM reasoning.It reaches a Codeforces Elo of 3020 and gold-medal-level performance in the 2025 IMO and CMO.Source: §4.1 · pp. 8, 11, Table 3 · p. 9, Table 4 · p. 10

4.2 Vision Task Evaluation

• Seed2.0 Pro delivers state-of-the-art vision results across mathematics, STEM, visual puzzles, perception, spatial understanding, and long-context document processing.Reported highlights include MathVision 88.8, LogicVista 81.4, DA-2K 92.3, and DUDE 72.4; Lite leads DynaMath with 70.5.Source: §4.2 · p. 12

4.3 Video Task Evaluation

• Seed2.0 pushes video understanding forward, achieving state-of-the-art results in motion perception, reasoning, streaming, long-video, and multi-video tasks.Seed2.0 Pro reaches 89.5 on VideoMME, 60.3 on CrossVid, and 37.4% accuracy on Morse-500; VideoCut further improves long-video benchmarks.Source: §4.3 · pp. 14–15

4.4 Fundamental Agentic Capacity

• Seed2.0 Pro ranks in the top tier for agentic capabilities, with particular strengths in search, deep research, vision agents, and structured artifact manipulation.Seed2.0 Lite ranks first or second on efficient agentic benchmarks while offering lower inference cost for latency-sensitive deployment.Source: §4.4 · pp. 16–17, Table 11 · p. 16, Table 12 · p. 17

4.5 Advanced Economically and Scientifically Valuable Tasks

• Seed2.0 Pro performs strongly on scientific discovery and economically valuable workflows, while Lite offers competitive advanced-task performance at favorable latency and cost.Pro leads AInstein Bench and XPert Bench, whereas it still trails strong baselines on DeR2 Bench and NL2Repo-Bench.Source: §4.5 · pp. 17–18, Table 13 · p. 18

5.1.1 Complex Code Generation

• Seed2.0 solved the TerminalBench 2.0 cryptanalysis task in 12 interaction rounds using structural analysis, meet-in-the-middle search, and layered validation.It validated the recovered key against 32 plaintext–ciphertext pairs, cross-verified the decrypt program, and decrypted 100 additional ciphertexts.Source: §5.1.1 · pp. 21–22

5.1.2 Project Repository Construction

• Seed2.0 constructed an installable Decouple-style Python repository from requirements alone in 37 interaction rounds, with all 22 model-written tests passing.The workflow covered specification analysis, implementation, testing, debugging, example execution, editable installation, and cross-directory import validation.Source: §5.1.2 · p. 22

5.1.3 Iterative Code Debugging

• Seed2.0 refactored Qt warning filtering while preserving runtime semantics, public APIs, and downstream compatibility, achieving 56 passing unit tests and verified import compatibility.The process used boundary scanning, compatibility wiring, regression testing, integration verification, and final cleanup.Source: §5.1.3 · pp. 22–23

5.1.4 Complex Iterative Debugging

• Across 79 turns, Seed2.0 upgraded DVC through exploration, implementation, and verification, handling cross-cutting OS, logging, CLI/API, concurrency, and output-behavior changes.The most involved change threaded --all-commits through command entry points, repository operations, revision traversal, and initialization while repeated tests provided acceptance feedback.Source: §5.1.4 · p. 25

5.1.5 Competition-Level Programming

• 73.02% Pass@8 earned Seed2.0 Pro Gold Medals in all five ICPC contests and significantly outperformed GPT-5.2 and Gemini-3-Pro.The evaluation used official test cases, time limits, and memory limits, with up to eight independent generations per problem.Source: §5.1.5 · p. 26

5.2.1 FreeCAD Operation

• In FreeCAD, Seed2.0 tracked stateful modeling context and recovered from misclicks, mode errors, dialogs, and UI lag through self-reflective corrective micro-plans.Its online control loop restored operational context and resumed the parametric workflow rather than merely documenting failures.Source: §5.2.1 · p. 26, §5.2 · p. 26
• Seed2.0 validates FreeCAD operations through redundant GUI and Python checks of final volume and area, reducing silent selection or document-state failures.The workflow checks Properties-panel values and independent Python readouts.Source: §5.2.1 · p. 28

5.2.2 CapCut Operation

• Seed2.0 executes frame-accurate CapCut workflows across splitting, transitions, audio, and effects, recovering from UI errors through reflective self-correction and timeline re-anchoring.It restores canonical invariants at 00:00:00:00 and 00:00:02:29 before continuing, preventing error propagation.Source: §5.2.2 · p. 28

5.2.3 Image-to-Code Generation

• Seed2.0 recreates image-based websites with specified animations and generates matplotlib 3D plots whose phase-space projection follows the input.The website includes rolling counts, XP bars, 3D tilt and highlight effects, and dynamic shimmer textures.Source: §5.2.3 · pp. 30–31

5.3 Multidisciplinary Scientific Research

• The multidisciplinary research cases evaluate Seed2.0 through scientific code generation and cross-domain analytical reasoning spanning quantum computing, numerical relativity, and computational chemistry.These tasks couple mathematical representations, physical conventions, and large codebases, creating subtle, high-cost errors.Source: §5.3 · p. 31, §5.3.1 · pp. 32, 34

5.3.1 Multidisciplinary Scientific Research Coding

• On Qiskit’s Solovay–Kitaev compiler, Seed2.0 diagnoses a SU(2)/SO(3) phase mismatch causing Pauli-Y to become −Y and localizes the fix to post-processing.It reproduces the pure π phase discrepancy, preserves the correct recursion, aligns operators by computing ϕ, and validates generalization across single-qubit gates.Source: §5.3.1 · pp. 33–34
• On Einstein Toolkit numerical-relativity code, Seed2.0 finds the missing AHFinder_dis.F routine and implements proper-distance computation using the full spatial metric and robustness checks.It traces the routine’s role in multi-horizon binary-black-hole simulations and distinguishes proper distance from coordinate separation.Source: §5.3.1 · pp. 35–36
• The computational-chemistry case involves a PySCF density-fitting bug producing an approximately 0.9 Hartree energy error for complex molecular-orbital coefficients.Resolving it requires decomposing complex density matrices into real and imaginary parts because low-level transformations support only real arrays.Source: §5.3.1 · p. 37
• Seed2.0 diagnosed a complex-valued density-fitting defect in PySCF, localized corruption to the K matrix, and produced a validated fix preserving the real-valued fast path.The fix decomposes complex density matrices into real and imaginary parts, applies the optimized real pathway twice, and recombines the results; post-patch energies matched references to approximately 10^-11 Hartree.Source: §5.3.1 · pp. 38, 40

5.4 Multidisciplinary Scientific Research Analysis

• Across the scientific-analysis cases, evaluation focused on technical soundness, internal consistency, domain conventions, justified methodological choices, reproducibility, and downstream transferability.The cases span theoretical derivations, computational modeling, experimental design, biomolecular simulation, and polymer chemistry.Source: §5.3.1 · p. 38, §5.4 · pp. 40, 42
• Seed2.0 generated proofs for two Erdős Problems, with human experts verifying correctness and the Erdős 1051 proof formalized using Seed-Prover 1.5.The case demonstrates systematic reasoning and rigorous proof construction for frontier-level mathematical research challenges.Source: §5.4 · p. 40
• For CBD–α7 nAChR analysis, Seed2.0 produced a logically structured CG-MD workflow with appropriate structure selection, validation metrics, GROMACS preprocessing, and downstream cross-validation strategies, but hallucinated PDB and literature details.The workflow serves as a useful initial protocol template, yet the inaccuracies require systematic manual verification before practical use.Source: §5.4 · p. 42
• Seed2.0 correctly reasoned through maleimide polyacetylene synthesis, including Schrock-catalyzed ROMP, the TEA/oxidant deprotonation–oxidation mechanism, and advantages over conventional copolymerization.It recognized that TEA alone is too weak to drive the equilibrium and that oxidation generates radical intermediates and shifts the transformation forward.Source: §5.4 · pp. 43–44
• In polymer chemistry, Seed2.0 correctly links oxidation to a planar, conjugated mPA backbone and explains how maleimide lowers LUMO energy for an air-stable n-type semiconductor.It also connects precursor sp3-hybridized carbon centers with backbone flexibility, solubility, and processability.Source: §5.4 · p. 44
• The scientific analyses have limitations: Seed2.0 misses the reaction’s charge-transfer complex mechanism and its Giantin tag choice may increase nonspecific binding or perturb Golgi architecture.Its mechanistic explanation therefore lacks depth in the key initiation and kinetic details.Source: §5.4 · pp. 44, 46
• Its Golgi protocol includes CRISPR/Cas9 targeting, contamination controls, quantitative purity assessment, and differential-testing criteria of FDR < 0.05 and |LFC| ≥1.2.The workflow specifies a Rosa26 reporter strategy and uses marker-based controls plus a PCC-based purity metric.Source: §5.4 · pp. 45–46
• Seed2.0 produces a feasible, scientifically sound Golgi workflow integrating cell-type-specific mouse engineering, Golgi isolation, and proteomics, lipidomics, and metabolomics.The strategy is designed to identify dysregulated Golgi molecules by cell type and brain region in a disease mouse model.Source: §5.4 · pp. 45–46

5.5 Automated Model-on-Model Behavioral Diagnostics

• The automated model-on-model diagnostic pipeline analyzes heterogeneous benchmark outcomes at scale and exposes quantified weaknesses, strategy trade-offs, and behavioral issues for model iteration.Its three-layer design combines data processing, scenario-adaptive analysis, and report synthesis across scores, behavioral statistics, outputs, traces, and execution trajectories.Source: §5.5 · p. 46

6 Conclusion

• Seed2.0 builds a reliable, forward-looking evaluation system from benchmarks grounded in users’ needs and complex real-world scenarios.This evaluation system guides the model’s development toward complex real-world tasks.Source: §6 · p. 46
• Seed2.0 targets long-tail knowledge and complex instruction following to improve reliability on complex, long-horizon real-world tasks.The model series takes a key step toward solving intricate real-world tasks.Source: §6 · p. 46
• Seed2.0 delivers world-leading reasoning intelligence, visual understanding, and search capabilities that address common needs for a broad user base.The model card documents real use cases showing initial ability to handle complex real-world tasks and bring greater value to hundreds of millions of users.Source: §6 · p. 48

7 Contributors

• Contributors are listed alphabetically by first name.The passage explicitly states the ordering convention for the author list.Source: §7 · p. 58
• Some listed names are internal aliases used within the company.The section qualifies the contributor list by explaining that some names refer to company-internal aliases.Source: §7 · p. 58
• The contributor list begins with names including Allan Jie, Anwen Hu, Aoyan Li, Astor Wu, and Baihan Shu.These names appear at the beginning of the supplied alphabetical list.Source: §7 · p. 58
• The list continues through contributors whose names begin with Cheng, Chen, and related entries.The supplied passages enumerate contributors across the Cheng and Chen name ranges.Source: §7 · p. 58
• The middle portion of the list includes extensive contributor ranges beginning with Feng, Gu, Han, Hong, Hua, and Jian.The supplied passages enumerate these consecutive alphabetical ranges.Source: §7 · p. 59
• The list extends through later alphabetical ranges from Jing and Jun to Zhi, Ziqiang, Ziyun, and Zuquan.The remaining passages enumerate contributors across the later J-, Z-, and related name ranges.Source: §7 · pp. 60–62

B.1 Task Overview

• The broader workflow requires recovering the round keys and decrypting all target ciphertexts into plaintexts.txt.The supplied task includes feal.c, decrypt.c, pairs.txt, and ciphertexts.txt as the attack inputs and implementations.Source: §B.1 · p. 70
• The attack targets four round keys derived from 20-bit seeds, while brute-forcing the total 2^80 keyspace is infeasible.The task uses 32 known plaintext-ciphertext pairs to recover k0, k1, k2, and k3 and decrypt 100 ciphertexts.Source: §B.1 · p. 70

Phase 1: Understanding the Cipher Structure

• Each 20-bit seed is expanded into a 32-bit round key, and encryption uses a four-round Feistel structure with pre- and post-round mixing.The F-function applies four byte transformations through the G-function.Source: Phase 1: Understanding the Cipher… · p. 71

Phase 2: Attack Design — Meet-in-the-Middle

• The inverse is required to work backward from ciphertext during the meet-in-the-middle attack.The attack computes forward states after round 0, backward states before round 3, and matches aligned intermediate states.Source: Phase 2: Attack Design —… · p. 71
• Meet-in-the-middle reduces the attack complexity from O(2^80) to O(2^21) by matching forward k0 and backward k3 intermediate states.Additional known plaintext-ciphertext pairs are then used to recover k1 and k2.Source: §B.1 · p. 70, Phase 2: Attack Design —… · p. 71

Phase 3: Implementing the F-function Inverse

• The inverse reverses the F-function’s G-transformations using a 2-bit right rotation and modular subtraction to recover intermediate byte values.The implementation defines rotr2 as the inverse of the original 2-bit rotation and reverses the G operations in sequence.Source: Phase 3: Implementing the F-function… · p. 72
• The inverse reconstructs the original bytes by undoing the G-function equations with XOR, subtraction modulo 256, and byte recombination.The final implementation returns the recovered bytes as a combined value.Source: Phase 3: Implementing the F-function… · p. 72
• F^-1(F(x)) = x passed for multiple test vectors, validating the implemented F-function inverse.The verification reported that all tested vectors passed.Source: Phase 3: Implementing the F-function… · p. 72

Step A: Build k0 Delta Map

• Because expand() is not injective, the attack precomputes a reverse lookup over 2^20 inputs, producing 1 ,047 ,882 entries with some collisions.The resulting exp_map stores each expanded value's corresponding k.Source: Phase 4: Key Recovery Implementation · p. 73
• The implementation builds a delta map with 933,294 unique deltas from intermediate values computed for candidate k0 values.Each entry maps ∆= A0 ⊕A1 to (k0, A0, A1).Source: Phase 4: Key Recovery Implementation · p. 73

Step B: Match with k3

• For Step B, the attack iterates k3 ∈[0, 220) and computes backward from the ciphertext to search for matches.When a match is found, it is used to derive the potential key material for subsequent verification.Source: Phase 4: Key Recovery Implementation · p. 73
• The attack finds candidate keys by matching k3-derived values, recovering K2 and k2, deriving k1, and checking K1 consistency.A match yields K2 = A0 ⊕B0, which is verified through exp_map before pair 2 derives k1 and pair 3 confirms consistency.Source: Phase 4: Key Recovery Implementation · p. 73

Key Verification Against All Pairs

• All 32 pairs verified successfully, confirming that the recovered keys are correct.Cross-verification with the provided ./decrypt binary matched the expected plaintext 367472469947502207.Source: Phase 5: Verification and Decryption · p. 74

Decrypting Target Ciphertexts

• The decryption reverses the final swap and processes the four Feistel rounds backward using the recovered round keys.The implementation restores the halves, iterates round numbers from 3 through 0, and applies the mixing function with each corresponding key.Source: Phase 5: Verification and Decryption · p. 74
• 100 ciphertext values were decrypted successfully and saved to plaintexts.txt.Each ciphertext was decrypted using the recovered keys.Source: Phase 5: Verification and Decryption · p. 74

B.3 Attack Performance Analysis

• ∼1018 times faster: the meet-in-the-middle attack requires 2 × 220 ≈2.1 × 106 operations versus 280 ≈1.2 × 1024 for brute force.The meet-in-the-middle strategy reduces complexity to O(221) by exploiting the Feistel structure.Source: §B.3 · p. 75, §B.4 · p. 75

B.4 Key Findings and Agent Behaviors

• The agent correctly identified direct brute force over the 280 keyspace as infeasible and autonomously designed a meet-in-the-middle strategy.The strategy exploits the cipher’s Feistel structure to reduce the search complexity to O(221).Source: §B.4 · p. 75
• The agent derived the inverse F-function by analyzing composed G-functions, rotations, modular arithmetic, and bit-rotation properties.This mathematical inversion was required to support the cryptanalytic attack.Source: §B.4 · p. 75
• The agent used multi-stage verification instead of accepting the first candidate key, including pair-based derivation, consistency checks, and validation against all 32 known pairs.Pairs 0 and 1 built the delta map, pair 2 derived k1, and pair 3 checked k1 consistency.Source: §B.4 · p. 75
• The recovered result was cross-checked with the original decrypt binary, strengthening confidence in the cryptanalysis.This binary cross-check supplemented the known-pair verification process.Source: §B.4 · p. 75
• The agent built reverse lookup tables, including expand and delta maps, to avoid redundant calculations during the search phase.This precomputation demonstrated awareness of search-time optimization.Source: §B.4 · p. 75
• The implementation used clean, modular helper functions that mirrored the original C implementation for consistency.Helpers included getleft, getright, merge, g_function, f_function, and f_inv_function.Source: §B.4 · p. 75

B.5 Recovered Keys and Final Results

• 100% accuracy was achieved when verifying all 4 recovered round keys against 32 known plaintext-ciphertext pairs, enabling decryption of 100 ciphertexts.The recovered 20-bit round key seeds are reported in Table 18.Source: §B.5 · p. 75, Table 18 · p. 75
• The decrypted plaintexts were saved to plaintexts.txt and cross-validated with the original decrypt binary.Source: §B.5 · p. 76

B.6 Lessons for Cryptanalysis Agent Design

• The agent analyzed source code to identify Feistel-network weaknesses and a reduced keyspace, then derived inverse functions algebraically for ciphertext recovery.Source: §B.6 · p. 76
• The agent selected meet-in-the-middle attacks from problem constraints and used precomputation with lookup tables to improve computational efficiency.Source: §B.6 · p. 76
• Progressive multi-stage validation reduced false-positive risk, while faithful C-to-Python translation preserved algorithmic correctness.Source: §B.6 · p. 76
• Reducing the key size from 128 bits to 20 bits made the cipher vulnerable to known-plaintext attacks within seconds on commodity hardware, underscoring the need for secure production key lengths.Source: §B.6 · p. 76

C.1 Task Overview

• The task required implementing a repository-level Python configuration library from a 39KB specification, including a complete project structure and 100% test coverage.The requested API included imports such as AutoConfig, Csv, and Choices from decouple.Source: §C.1 · p. 76
• The library supports environment variables, .env files, .ini files, type conversion, validation, Docker secrets, defaults, and custom casters.Type conversion requirements included booleans, integers, CSV values, and custom casting; validation included Choices and default values.Source: §C.1 · p. 76

C.2 Agent Execution Trace

• The agent systematically parsed the specification, extracting exact interfaces, file requirements, conversion rules, quote handling, and documentation-embedded test cases.Code snippets in the specification were treated as implicit behavioral tests.Source: §C.2 · p. 76
• Implementation decisions addressed Python 3.12 ConfigParser compatibility, matching-quote stripping, graceful missing Docker-secret directories, and empty-string boolean conversion.The agent replaced unavailable read_config() with read_file() and added an early return so empty strings become False rather than raising ValueError.Source: §C.2 · p. 77
• All 22 tests passed after fixing empty-string boolean conversion and correcting fixture scope, followed by verification of package installability.The initial test run exposed one failure, and the subsequent run exposed four fixture-scope errors before the final passing run.Source: §C.2 · p. 77

C.3 Key Agent Capabilities Demonstrated

• The agent produced a complete installable package with API-compatible modules, setup metadata, documentation, licensing, and PyPI distribution infrastructure.The implementation covered more than 300 lines and included RepositoryEnv, RepositoryIni, RepositorySecret, Config, AutoConfig, Csv, and Choices classes.Source: §C.3 · p. 77, §C.2 · p. 77
• The agent demonstrated specification parsing, error-driven development, backwards-compatible API exports, and production-quality package generation.It debugged an ImportError using the Python 3.12 API documentation and generated LICENSE, MANI-FEST.in, and setup.cfg files.Source: §C.3 · p. 77

C.4 Performance Metrics

• Task Completed: Fully functional, installable Python package matching all specifications.Installation and configuration checks also succeeded.Source: §C.4 · p. 78

D.1 Task Overview

• The SWE-bench Pro task relocates Qt logging utilities and tests while preserving filtering behavior, existing imports, and warning-suppression functionality.The requested change moves hide_qt_warning and QtWarningFilter from log.py to qtlog.py and relocates tests to test_qtlog.py.Source: §D.1 · p. 79

D.2 Agent Execution Trace

• The migration required coordinated changes across qtlog.py, log.py, tests, and dependent usage to maintain consistency.Repository exploration identified the relevant modules, tests, and qtnetworkdownloads.py as a caller of log.hide_qt_warning().Source: §D.2 · p. 79
• The refactored QtWarningFilter preserves exact whitespace handling through record.msg.strip().startswith(), enabling matches with leading or trailing spaces.hide_qt_warning adds the filter to the selected logger and removes it afterward.Source: §D.2 · p. 80
• Re-exporting hide_qt_warning and QtWarningFilter from log.py makes both legacy and new import paths resolve to the same function object.The original implementation was removed from log.py while compatibility imports were retained.Source: §D.2 · pp. 80–81
• 51 tests passed in test_log.py and 5 tests passed in test_qtlog.py after the Qt logging refactor.The migrated tests cover nonmatching patterns, exact and prefix suppression, and whitespace handling.Source: §D.2 · p. 81
• The real qtnetworkdownloads.py usage pattern filters QNetworkReplyImplPrivate warnings while allowing Other warning messages to pass through.This confirms the refactored context manager behaves correctly for an existing caller.Source: §D.2 · p. 81

D.3 Refactoring Analysis

• The code migration is summarized in Table 20 under the refactoring analysis.Source: Table 20 · p. 81

D.4 Key Software Engineering Practices

• The agent performed refactoring in five systematic steps rather than bulk copy-paste, reducing the risk of breaking changes.Source: §D.4 · p. 81
• A Python re-export pattern, from qtlog import *, preserved the existing API surface while reorganizing internals.Source: §D.4 · p. 82
• Tests were moved after the code and all tests were validated before success was declared, guarding against behavior regression.Source: §D.4 · p. 82
• The agent searched the codebase for all usages of hide_qt_warning to identify affected modules before changing the implementation.Source: §D.4 · p. 82
• The resulting refactoring caused zero breaking changes and improved code organization.Source: §D.4 · p. 82
• All 56 tests passed after the migration, supporting the absence of behavior regression.Source: §D.4 · p. 82

E.1 Natural Language Proving

• Seed2.0 Pro adopts an iterative solve-verify-refine pipeline [52] for natural-language mathematical proving.The pipeline is designed around generating, checking, and refining solutions.Source: §E.1 · p. 82
• The model generates candidate solutions, autonomously identifies logical flaws, and refines outputs into systematically reasoned, mathematically rigorous proofs.Table 2 reports the results, while the described process validates deduction step by step rather than relying on specious correct final answers.Source: §E.1 · p. 82

E.2 Formal Theorem Proving

• 35.5% Pass@8 makes Seed2.0 Pro state-of-the-art among comparable general-purpose models, surpassing Gemini-3-Pro and Seed-1.5-Prover; Seed2.0 Lite reaches 30.5%.The evaluation uses 200 formalized Putnam Competition problems with Lean and Python tool access in an agent-based multi-turn setup.Source: §E.2 · p. 82

E.3 Case study on Erdos Problems

• The Erdős case studies cover combinatorial geometry in Erdős 652 and analytic number theory in Erdős 1051.They are presented as demonstrations of the AI agent’s capability to solve challenging open problems from Paul Erdős’s problem collection.Source: §E.3 · p. 82

E.3.1 Erdos 652: Distinct Distances in Planar Point Sets

• For Erdős 652, the construction builds a multigraph from point-circle incidences, with I = kn − k and E ≥ kn − k − kt.The graph uses circles centered at the first k points and connects consecutive points on each circle.Source: §E.3.1 · p. 83
• Bounding parallel edges gives R ≤ kt(k + 1), hence the simplified graph satisfies E′ ≥ kn − k − kt(k + 2).These bounds control the loss when converting the multigraph into a simple graph suitable for the crossing-number argument.Source: §E.3.1 · pp. 83–84

E.3.2 Erdos 1051: Irrationality of Infinite Series

• The proof establishes that S is irrational by contradiction: assuming S = P/Q for coprime positive integers yields x_m → 0, contradicting x_m ≥ 1/Q > 0 for every m ≥ 1.Source: §E.3.2 · pp. 85, 87
• The Erdős 1051 argument first derives a positive lower bound x_m ≥ 1/Q from rationality and then obtains a recurrence from a loose tail bound.The proof uses integer numerators after clearing denominators and the monotonicity of the sequence a_n.Source: §E.3.2 · p. 85
• Logarithmic normalization shows r_k is bounded and converges, while d_k converges to D ≥ log L > 0 under the liminf assumption.This positive limiting quantity enables the later geometric tail estimate.Source: §E.3.2 · p. 86
• The geometric tail bound yields x_m → 0, contradicting the rationality lower bound and completing the Erdős 1051 proof by contradiction.For sufficiently large indices, consecutive tail terms have ratio at most 1/2, so the tail is bounded by a geometric series.Source: §E.3.2 · p. 87