SC-ARENA: Multidisciplinary Evaluation Arenas

• SC-ARENA is a non-unique research designation used across various fields, encompassing evaluation benchmarks for LLM negotiation, single-cell analysis, and cosmic-ray detection.
• It standardizes methodologies such as phase-aware negotiation in stochastic games, human-preference evaluation for search-augmented models, and ontology-guided assessments in cellular biology.
• The framework inspires related constructs in gaming, social navigation, and multi-agent control, highlighting the need for context-specific interpretation in diverse technical literatures.

SC-ARENA is a non-unique research designation used across several unrelated technical literatures. In recent arXiv work, it refers to SidConArena, an evaluation framework for LLM agents in open-ended, positive-sum bargaining; Search Arena, a human-preference benchmark for search-augmented LLMs; SC-Arena, a natural-language benchmark for single-cell reasoning; and a South Pole radio surface-array concept for cosmic-ray air showers. Related papers also identify StarCraft II Battle Arena and Arena 3.0 as the relevant “arena” constructs in their domains while explicitly noting that those works do not formally use the name SC-ARENA (Feng et al., 24 Jun 2026, Miroyan et al., 5 Jun 2025, Zhao et al., 26 Feb 2026, Schröder, 2018, Li et al., 18 Dec 2025, Kästner et al., 2024).

1. Terminological scope

The literature does not treat SC-ARENA as a single canonical acronym. One paper states that, in its context, “SC-ARENA” refers to SidConArena and that the authors use “SidConArena” throughout, with “SC-ARENA” described as an acceptable abbreviation for discussion. Another introduces “SC-ARENA (Search Arena).” A third adopts the formal title “SC-Arena.” By contrast, the StarCraft paper states that it “does not use the name ‘SC-ARENA,’” and the Arena 3.0 paper likewise “does not explicitly use the term ‘SC-ARENA’.” A separate South Pole proceeding uses SC-ARENA for a radio surface-array concept (Feng et al., 24 Jun 2026, Miroyan et al., 5 Jun 2025, Zhao et al., 26 Feb 2026, Li et al., 18 Dec 2025, Kästner et al., 2024, Schröder, 2018).

Usage	Domain	Naming status
SidConArena	LLM bargaining benchmark	“SC-ARENA” acceptable abbreviation
Search Arena	Search-augmented LLM evaluation	“SC-ARENA (Search Arena)”
SC-Arena	Single-cell reasoning benchmark	Formal title
SC-ARENA	South Pole radio surface array	Concept name
SC2BA	Multi-agent StarCraft arena	Paper says it does not use “SC-ARENA”
Arena 3.0	Social navigation stack	Paper does not explicitly use “SC-ARENA”

This naming heterogeneity is structurally important. It means that references to SC-ARENA must be interpreted from disciplinary context rather than from the acronym alone.

2. SidConArena: open-ended, positive-sum bargaining

In the bargaining literature, SC-ARENA denotes SidConArena, a benchmark framework for evaluating LLM agents in open-ended, mixed-motive bargaining. It formalizes a multi-player economy as a finite-horizon partially observable stochastic game with three coupled phases: natural-language negotiation with binding trades, deterministic converter-based production, and sealed-bid auctions for long-term assets. The formal model is

$$\mathcal{G}=\langle \mathcal{N}, \mathcal{S}, \mathcal{A}, \mathcal{P}, H, V_{terminal}, \mathcal{O}, Z \rangle,$$

with terminal-value optimization rather than dense per-step rewards, and the agent objective is

$$J(\pi_i)=\mathbb{E}[V_{terminal}(s_{H,i})]=\mathbb{E}[\mathbf{w}\cdot \mathbf{r}_{H,i}].$$

Negotiation is implemented through natural-language messages plus binding trade vectors

$$\tau_{i\leftrightarrow j}=(\Delta\mathbf{r}_{i\to j},\;\Delta\mathbf{r}_{j\to i}),$$

validated and executed atomically once both parties confirm. Production uses converters $\mathcal{F}_{t,i}$ with deterministic state transitions, while the auction phase uses a Simultaneous Multi-Track Priority Auction with sealed bids in Ships. Partial observability is induced by private inventories, public board signals, phase-relevant market context, and sealed bids (Feng et al., 24 Jun 2026).

The agent interface is explicitly phase-aware. Observations are structured into private, public, market, and interaction channels, and a “Brain” dispatcher routes them to specialized modules: TurnPlanCaller, TradeCaller, EconomyCaller, BidCaller, PickCaller, and DiscardColonyCaller. A neural-symbolic action interface converts free-form reasoning into validated function-call actions, and asynchronous event-driven execution permits concurrent negotiation without sacrificing atomicity or reproducibility.

Evaluation is centered on terminal economic performance, with social welfare $W=\sum_i \mathbf{w}\cdot \mathbf{r}_{H,i}$ as a natural aggregate measure of positive-sum surplus creation. For heterogeneous settings, the framework also computes model-level Elo ratings from terminal outcomes on a 1500 baseline using zero-sum tournament aggregation. Empirically, frontier models such as GPT-5 and Gemini-3-Flash-Preview achieve higher terminal scores in homogeneous self-play, but the benchmark also exposes recurring pathologies: misvaluation of Ships, passive bargaining, and limited long-horizon investment planning. The reported interpretation is that optimizing local plausibility does not suffice when terminal scoring rewards investment trajectories.

SidConArena is distinctive because it couples open-ended negotiation, validated contracts, deterministic production, and imperfect-information auctions in a single POSG. This makes it a benchmark for mixed-motive coordination rather than only adversarial play or static reasoning.

3. Search Arena: human-preference evaluation of search-augmented LLMs

In the retrieval-augmented LLM literature, SC-ARENA refers to Search Arena, a large-scale, crowd-sourced, human-preference benchmark and dataset for studying models that query and read the live web at inference time. The released dataset contains 24,069 conversations and 12,652 paired human preference votes collected over seven weeks from 11,650 users in 136 countries. It spans 70+ languages, with approximately 58% English, approximately 11–12% Russian, approximately 7% Chinese, and more than 11% multilingual prompts. Multi-turn interaction is explicit: 22.4% of conversations have more than one turn. The benchmark defines nine intent categories—Factual Lookup, Information Synthesis, Analysis, Recommendation, Explanation, Creative Generation, Guidance, Text Processing, and Other—and Factual Lookup accounts for only 19.3% of queries (Miroyan et al., 5 Jun 2025).

Search Arena integrates search-enabled models as a “Search” tab in Chatbot Arena, where two anonymous models respond side by side and users may vote A, B, Tie, or Both-bad. The released records include anonymized model identities, full conversation history, reasoning traces, retrieved URLs, inline citations, and derived annotations. Pairwise preference modeling follows Bradley–Terry, with feature-controlled comparisons of the form

$$P(i \succ j)=\sigma\big((s_i-s_j)+\beta^\top(x_i-x_j)\big).$$

This framework supports Elo-scaled leaderboards as well as analyses of length, citation count, and source-type effects.

The central empirical result is a credibility gap between perceived and actual support. Users prefer longer answers overall, with $\beta_{length}=0.334$, and the effect is smaller for Factual Lookup, where $\beta_{length,factual}=0.156$. Users also prefer more citations, with $\beta_{citations}=0.209$. Source-type analyses show positive associations for tech/code platforms $(\beta_{tech}=0.073)$, community platforms/blogs $$J(\pi_i)=\mathbb{E}[V_{terminal}(s_{H,i})]=\mathbb{E}[\mathbf{w}\cdot \mathbf{r}_{H,i}].$$0, and social media $$J(\pi_i)=\mathbb{E}[V_{terminal}(s_{H,i})]=\mathbb{E}[\mathbf{w}\cdot \mathbf{r}_{H,i}].$$1, while Wikipedia has a negative association $$J(\pi_i)=\mathbb{E}[V_{terminal}(s_{H,i})]=\mathbb{E}[\mathbf{w}\cdot \mathbf{r}_{H,i}].$$2. Most notably, an attribution pipeline found that users reward both supporting citations $$J(\pi_i)=\mathbb{E}[V_{terminal}(s_{H,i})]=\mathbb{E}[\mathbf{w}\cdot \mathbf{r}_{H,i}].$$3 and irrelevant citations $$J(\pi_i)=\mathbb{E}[V_{terminal}(s_{H,i})]=\mathbb{E}[\mathbf{w}\cdot \mathbf{r}_{H,i}].$$4, whereas contradicting citations were not significant. The dataset therefore supports direct study of the gap between attribution quality and perceived credibility.

Cross-arena experiments further separate search-conditioned and non-search-conditioned settings. In Text Arena, 544 battles comparing Gemini-2.5 Pro Experimental with and without web search yielded 26% search-preferred, 28% non-search-preferred, and 45% ties, with no significant aggregate difference $$J(\pi_i)=\mathbb{E}[V_{terminal}(s_{H,i})]=\mathbb{E}[\mathbf{w}\cdot \mathbf{r}_{H,i}].$$5. In Search Arena, 315 battles of a non-search model against search-enabled models yielded 40% search-preferred, 29% non-search-preferred, and 31% ties, with a significant difference $$J(\pi_i)=\mathbb{E}[V_{terminal}(s_{H,i})]=\mathbb{E}[\mathbf{w}\cdot \mathbf{r}_{H,i}].$$6. The paper’s conclusion is that web search does not degrade overall chat performance in general settings, but reliance solely on parametric knowledge significantly harms performance when users expect grounded answers.

4. SC-Arena: single-cell reasoning and knowledge-augmented evaluation

In single-cell biology, SC-Arena is a natural-language benchmark and judging framework tailored to single-cell foundation models. Its core abstraction is the “Virtual Cell,” which jointly represents intrinsic attributes and gene-level interactions. The benchmark defines five natural-language tasks as typed mappings: Cell Type Annotation $$J(\pi_i)=\mathbb{E}[V_{terminal}(s_{H,i})]=\mathbb{E}[\mathbf{w}\cdot \mathbf{r}_{H,i}].$$7, Cell Captioning $$J(\pi_i)=\mathbb{E}[V_{terminal}(s_{H,i})]=\mathbb{E}[\mathbf{w}\cdot \mathbf{r}_{H,i}].$$8, Cell Generation $$J(\pi_i)=\mathbb{E}[V_{terminal}(s_{H,i})]=\mathbb{E}[\mathbf{w}\cdot \mathbf{r}_{H,i}].$$9, Perturbation Prediction $$\tau_{i\leftrightarrow j}=(\Delta\mathbf{r}_{i\to j},\;\Delta\mathbf{r}_{j\to i}),$$0, and Scientific QA $$\tau_{i\leftrightarrow j}=(\Delta\mathbf{r}_{i\to j},\;\Delta\mathbf{r}_{j\to i}),$$1. The Cell Ontology is modeled as a directed acyclic graph $$\tau_{i\leftrightarrow j}=(\Delta\mathbf{r}_{i\to j},\;\Delta\mathbf{r}_{j\to i}),$$2, with shortest-path distance $$\tau_{i\leftrightarrow j}=(\Delta\mathbf{r}_{i\to j},\;\Delta\mathbf{r}_{j\to i}),$$3 used to interpret semantic proximity between ontology labels (Zhao et al., 26 Feb 2026).

The evaluation framework is explicitly knowledge-augmented. Each judged instance is

$$\tau_{i\leftrightarrow j}=(\Delta\mathbf{r}_{i\to j},\;\Delta\mathbf{r}_{j\to i}),$$4

where $$\tau_{i\leftrightarrow j}=(\Delta\mathbf{r}_{i\to j},\;\Delta\mathbf{r}_{j\to i}),$$5 is the prompt, $$\tau_{i\leftrightarrow j}=(\Delta\mathbf{r}_{i\to j},\;\Delta\mathbf{r}_{j\to i}),$$6 the model response, $$\tau_{i\leftrightarrow j}=(\Delta\mathbf{r}_{i\to j},\;\Delta\mathbf{r}_{j\to i}),$$7 retrieved external knowledge, and $$\tau_{i\leftrightarrow j}=(\Delta\mathbf{r}_{i\to j},\;\Delta\mathbf{r}_{j\to i}),$$8 the gold reference. An evaluator LLM assigns a discrete score $$\tau_{i\leftrightarrow j}=(\Delta\mathbf{r}_{i\to j},\;\Delta\mathbf{r}_{j\to i}),$$9, rescaled as

$\mathcal{F}_{t,i}$0

Knowledge grounding is task-specific: Cell Ontology identifiers and definitions for annotation, CellMarker for cell generation, NCBI/UniProt/GO for perturbation prediction, and PubMed articles for scientific QA. The stated purpose is to overcome the brittleness of exact match, BLEU, and ROUGE by rewarding ontology-aware near-misses, penalizing biologically implausible assertions, and generating evidence-grounded rationales.

The benchmark draws on 608 representative profiles from CZ CELLxGENE Discover for Cell Type Annotation, Cell Captioning, and Cell Generation; 138 genetic interventions from Norman (2019) and Adamson (2016) for Perturbation Prediction; and 254 questions from 100 PubMed papers for Scientific QA. Experiments include general-purpose models such as Qwen2.5, Qwen3, GPT-4o, DeepSeek-R1, and Kimi-K2, together with domain-specialized systems such as C2S-Pythia-410M, scGPT, scGenePT variants, and Cell-o1.

The reported results show no model reaching a “reliable virtual cell.” The best overall general models are Kimi-K2 with a total of 277.16 and DeepSeek-R1 with 276.66, both below a nominal passing score of 300. Task bests among general models are 40.81 for Cell Type Annotation, 67.89 for Cell Captioning, 63.04 for Cell Generation, 37.54 for Perturbation Prediction, and 74.48 for Scientific QA. A domain-specific highlight is C2S-Pythia-410M scoring 47.34 on Cell Type Annotation, above much larger general LLMs. The evaluation framework also reports strong biological alignment: evaluator scores correlate with negative ontology distance in annotation $\mathcal{F}_{t,i}$1, percentage of correct marker genes in generation $\mathcal{F}_{t,i}$2, and DEG cosine similarity in perturbation prediction $\mathcal{F}_{t,i}$3, while rankings for captioning and scientific QA align with expert preferences $\mathcal{F}_{t,i}$4 and $\mathcal{F}_{t,i}$5. The paper characterizes the resulting performance pattern as “fluent but not faithful.”

5. Related arena usages in multi-agent control and social navigation

A separate line of work in multi-agent reinforcement learning introduces the “StarCraft II Battle Arena” (SC2BA) rather than SC-ARENA. The paper explicitly states that it does not use the name SC-ARENA, and that if one is looking for a StarCraft “arena,” SC2BA is the arena introduced by that work. SC2BA replaces the static built-in bots of SMAC with algorithm-vs-algorithm adversaries, and it is organized around three design principles: fairness, usability, and customizability. Fairness is implemented through matched combat forces in symmetric cases, symmetric spatial layouts, mirrored map perspective, fixed sight and attack ranges, and standardized interaction. Usability includes text-based scenario configuration, a unified map, a Gym-like interaction API, Linux headless SC2 execution, a PySC2 wrapper, and the APyMARL PyTorch library. Customizability includes arbitrary red/blue algorithm assignment, randomized or pre-trained starting models, symmetric and asymmetric scenarios, and multiple adversarial paradigms. Benchmarking uses dual-algorithm paired adversary and multi-algorithm mixed adversary modes, with primary evaluation by median win rate over 32 evaluation episodes and five independent runs. The reported findings include win-rate fluctuations as the norm, no universal dominator, strong sensitivity to scenario complexity, and marked sensitivity to troop asymmetry (Li et al., 18 Dec 2025).

Arena 3.0 occupies a different domain: social navigation in collaborative and highly dynamic environments. The paper does not explicitly use the term SC-ARENA, but the accompanying interpretation describes it as the socially compliant, socially aware instantiation of the Arena platform. The software stack unifies Arena-core, Pedsim ROS, a Pedsim Waypoint Plugin, semantic layers, an MBF-based navigation suite, task generation, and benchmarking across Flatland, Gazebo, and Unity. It integrates multiple social force models, including Helbing–Molnár–Farkas–Vicsek, PySocialForce, DeepSocial, evacuation and bonding variants, and ORCA/RVO2. The benchmark suite reports standard motion metrics such as success rate, collision rate, time-to-goal, path length, jerk, and curvature, together with social proxies such as time in private zone, time facing pedestrians, and time seen by pedestrians. In the reported benchmark demonstration, the LfLH planner shows stronger social metrics but lower efficiency, whereas the ROSNav RL planner is efficient but spends longer in private zones on average, approximately 5.6 seconds (Kästner et al., 2024).

Taken together, these two usages show that “arena” can denote either an adversarial evaluation substrate for MARL or a simulator-agnostic stack for socially compliant navigation, even when SC-ARENA is not the formal system name.

6. South Pole radio surface-array usage and IceCube-Gen2 context

In astroparticle physics, SC-ARENA denotes a radio surface-array concept co-deployed on the IceCube/IceTop site at the South Pole. Its stated motivation is to add a continuous, calorimetric measurement of the electromagnetic air-shower component and sensitivity to the depth of shower maximum, $\mathcal{F}_{t,i}$6. The South Pole is described as especially suitable because thermal and Galactic noise dominate, persistent human-made RFI is limited, and temporary RFI can be rejected with externally triggered radio readout. The proposed array uses wideband antennas optimized around 100–190 MHz, with SKALA identified as a candidate antenna and $\mathcal{F}_{t,i}$7 assumed in CoREAS simulations. For vertical showers, the radio footprint has diameter of order 250 m, and full coverage across the approximately $\mathcal{F}_{t,i}$8 IceTop area requires on the order of 80 antennas. Using the criterion $\mathcal{F}_{t,i}$9 in at least three stations, simulations indicate a threshold of a few PeV. Expected performance figures are approximately 10–15% energy resolution, approximately $W=\sum_i \mathbf{w}\cdot \mathbf{r}_{H,i}$0 $W=\sum_i \mathbf{w}\cdot \mathbf{r}_{H,i}$1 resolution, direction resolution $W=\sum_i \mathbf{w}\cdot \mathbf{r}_{H,i}$2, and zenith coverage from $W=\sum_i \mathbf{w}\cdot \mathbf{r}_{H,i}$3 to approximately $W=\sum_i \mathbf{w}\cdot \mathbf{r}_{H,i}$4 (Schröder, 2018).

The physics case is multi-component. Radio provides an electromagnetic calorimetric channel unaffected by snow accumulation, while IceTop ice-Cherenkov tanks, scintillators, and in-ice photomultipliers provide complementary measurements of electrons, muons, and high-energy muons. The intended applications include improved mass-composition measurements, mass-dependent anisotropy searches, tests of hadronic interaction models, inclined-shower measurements, and PeV photon searches from the Galactic Center, which is continuously visible from the South Pole at zenith approximately $W=\sum_i \mathbf{w}\cdot \mathbf{r}_{H,i}$5.

A later proceeding on the IceCube-Gen2 Surface Array describes a related hybrid architecture with elevated radio antennas and scintillation detectors above the footprint of the enlarged Optical Array. That proceeding specifies about 160 surface stations at approximately 240 m spacing, with eight scintillation panels and three elevated SKALA v2 radio antennas per station, a 50–350 MHz radio band, and about 50–60 antennas per $W=\sum_i \mathbf{w}\cdot \mathbf{r}_{H,i}$6. Radio is externally triggered rather than self-triggered, using scintillator coincidence, periodic software triggers, and deep Optical Array triggers on high-energy muon tracks. The coincident aperture for air-shower events detected by both surface and deep arrays is expected to increase by about a factor of 30 relative to current IceCube with IceTop. Prototype measurements indicate a radio threshold around 30 PeV at present, with simulations predicting full efficiency for a substantial fraction of arrival directions beginning at a few $W=\sum_i \mathbf{w}\cdot \mathbf{r}_{H,i}$7 PeV and the lowest full-efficiency threshold in the $W=\sum_i \mathbf{w}\cdot \mathbf{r}_{H,i}$8 zenith range (Schröder, 2023).

Within this South Pole usage, SC-ARENA denotes not a software benchmark but a detector concept: a radio-enhanced surface array designed to add electromagnetic calorimetry, $W=\sum_i \mathbf{w}\cdot \mathbf{r}_{H,i}$9 sensitivity, and inclined-shower acceptance to the IceCube program.