ARM2: Forward Calorimetry & Adaptive AI

• For hadronic physics, the paper demonstrates that LHCf Arm2’s tungsten-scintillator calorimeter precisely measures very-forward photons, neutrons, and mesons to constrain cosmic-ray interaction models.
• The detector’s performance metrics show energy-dependent efficiency (rising from ~52% to ~72%) and refined position resolution, though discrepancies with existing hadronic models remain.
• ARM2 AI is an adaptive multimodal reasoning framework that integrates vision tokens and executable code, improving accuracy by 8–10% while reducing token usage by over 70%.

ARM2 refers to two distinct but prominent entities within contemporary scientific literature: the LHCf Arm2 forward calorimeter, an instrumental component in very-forward physics at high-energy colliders, and ARM2, the Adaptive Reasoning Model with Vision Understanding and Executable Code in AI research. This article delineates both, providing rigorous coverage of each within their respective domains—experimental hadronic physics and AI-driven multimodal reasoning.

1. LHCf Arm2: High-Energy Forward Calorimetry and Measurements

The LHCf Arm2 detector is a tungsten-scintillator sampling calorimeter positioned 141.05 m from the ATLAS IP1, dedicated to measuring neutral particle (γ, n, π⁰, η, etc.) production at very high pseudorapidity ($\eta > 8.4$) in hadron collisions at the LHC. Arm2 comprises two towers (25×25 mm² and 32×32 mm²), each with 22 tungsten absorber plates (7 mm) interleaved with 16 Gd₂SiO₅ (GSO) scintillator layers (1 mm), totaling approximately 21 cm (44 $X_0$ or 1.6 $\lambda_I$) longitudinal depth. X–Y silicon microstrip planes with 160 μm pitch, embedded at four depths, provide transverse shower imaging for precise impact and particle-type discrimination (Adriani et al., 2018, Adriani et al., 2020).

Key performance metrics include a hadron detection efficiency rising from ≃52% at 1 TeV to ≃72% at 6.5 TeV, an energy resolution degrading from ≃28% to ≃38% across the same range, and a position resolution improving from ≃300 μm (1 TeV) to ≃100 μm (>6 TeV). Absolute energy-scale uncertainty is controlled to ±4.5%, with dominant systematics from energy scale, PID purity, and unfolding/model-dependence of corrections (Adriani et al., 2018, Adriani et al., 2020).

2. Arm2 in Forward Particle Production and Hadronic Model Constraints

Arm2’s core physics mission is the measurement of inclusive forward neutron, photon, π⁰, and η meson production at the highest accessible $\eta$ and $x_F$, crucial for constraining hadronic interaction models employed in ultra-high-energy cosmic ray (UHECR) air-shower simulations (Piparo et al., 2023). At $\sqrt{s} = 13$ TeV in p–p collisions, Arm2 established the first measurements of forward neutron energy flow, differential cross sections, and inelasticity in six distinct $\eta$ bins, with differential neutron production $d^2\sigma_n/dE\,d\eta$ exhibiting peaks shifting from ≃5 TeV (very forward) to ≃1.5 TeV (lower $\eta$) (Adriani et al., 2020, Adriani et al., 2018). None of the post-LHC hadronic models (e.g., EPOS-LHC, QGSJET II-04, SIBYLL 2.3, DPMJET 3.06, PYTHIA 8.212) satisfactorily reproduce both the spectral shapes and normalizations, with discrepancies most pronounced at the most forward $\eta$.

The Arm2 program also provided the first direct measurement of $X_0$0 meson production (with $X_0$1 GeV/c, as a function of $X_0$2), offering critical constraints on strange-quark hadronization in the hadronic cascade. These results reveal that leading generators systematically overpredict the high-$X_0$3 yield except for QGSJET II-04, which shows closer, though still imperfect, agreement (Piparo et al., 2023).

Table: Summary of Arm2 Forward Hadronic Observables at 13 TeV (select values for illustration)

Observable	Value/Range	Systematic Error
$X_0$4 (avg inelasticity)	$X_0$5	QGSJET II-04: 0.533 agreement
$X_0$6 (energy flow, $X_0$7)	$X_0$8 GeV	Dominated by energy scale, unfolding
$X_0$9 (η mesons, $\lambda_I$0 GeV/c)	$\lambda_I$1	Total uncertainty includes acceptance, energy scale

These discrepancies directly impact UHECR air-shower predictions, particularly the forward baryon inelasticity and the muonic and electromagnetic shower components, underscoring the unique value of Arm2 data in reducing model uncertainties (Adriani et al., 2020, Piparo et al., 2023).

3. RHICf/Arm2: Transplantation and Spin Physics

Arm2 was also proposed as the core of the RHICf experiment at the Relativistic Heavy Ion Collider for precise measurements of very forward photons, neutrons, and π⁰s in $\lambda_I$2+$\lambda_I$3 collisions at $\lambda_I$4 GeV. Installed at 18 m from the PHENIX IP, Arm2 enables access to the $\lambda_I$5 region, testing Feynman scaling and providing critical cross-comparisons with LHCf results at fixed $\lambda_I$6 and $\lambda_I$7. The proposed RHICf program leverages Arm2’s high resolution for single-shower spectra and uniquely extends transverse single-spin asymmetry ($\lambda_I$8) measurements for forward neutrons into new $\lambda_I$9 domains, with projected statistical uncertainties $\eta$0 for $\eta$1 (Itow et al., 2014).

4. ARM2: Adaptive Reasoning in Large Multimodal Models

ARM2 (Adaptive Reasoning Model with Vision Understanding and Executable Code) is a unified AI reasoning framework that dynamically adapts its output format and length based on task complexity. ARM2 builds on a multimodal transformer backbone (e.g., Qwen-2.5-VL-7B), combining tokenized text and vision features (images embedded as $\eta$2 tensors) (Xie et al., 9 Oct 2025).

ARM2 introduces a learned format selection mechanism, allowing the model to stochastically choose among Direct Answer (DA), Short Chain-of-Thought (CoT), Long CoT, Code-Text, and Code-Exec (executed Python code) formats. The choice is controlled by a lightweight selector attached to the [CLS] token at the first decoding step, conditioning subsequent generation and reinforcement reward.

The reward objective is a composition of correct-answer reward, format encouragement (to avoid mode collapse), length-aware penalty (to mitigate “over-thinking” and keep token usage efficient), and a cosine decay schedule. The grouped policy optimization (GRPO) variant of PPO is used for policy updates, with explicit formulas for group-based reward amplification and exponential length penalty:

$\eta$3

Length penalty hyperparameter ($\eta$4) governs the accuracy/efficiency tradeoff; $\eta$5 is the default.

5. Integration of Vision and Code Execution in ARM2

ARM2 subsumes vision reasoning by embedding image tokens alongside text input, supporting both short-form rationales and code-based computations referencing visual content. In “Code-Exec” mode, ARM2 emits Python snippets, which are executed by an external interpreter; the result is injected back as an <OBSERVATION> token, supporting verified reasoning and computation. Empirically, executable code formats raise accuracy on out-of-distribution math/vision tasks by 8–10%, and reduce average token usage by 40–50% on such benchmarks.

6. Training Strategy, Evaluation, and Analysis of ARM2

ARM2 is trained in three stages: synthetic data construction (from AQuA-Rat, VisualWebInstruct, etc.), supervised fine-tuning (via LoRA adapters), and length-aware RL (GRPO-alp), using both text and vision-heavy benchmarks. Calibrated format and length penalties ensure adaptive selection of efficient reasoning strategies without accuracy loss.

In comparative evaluation, ARM2 preserves performance parity (in-domain accuracy 51.2%, out-of-domain 57.1%) with strong RL-trained baselines (Qwen-GRPO at 51.8%/57.1%), while slashing average output tokens by over 70%. Ablation analyses confirm that both format encouragement and code execution are critical; omitting either results in degraded accuracy or sharply increased inference cost.

7. Significance, Limitations, and Future Prospects

For hadronic physics, LHCf Arm2 represents a precision tool for forward particle measurement, providing data that invalidate key predictions of current models and force retuning of cosmic-ray simulators. For AI, ARM2 provides a viable multiformat adaptive reasoning paradigm that negotiates the Pareto frontier of accuracy and brevity, integrating code execution as a robust verification mechanism.

The dominant limitations for LHCf/Arm2 are model-dependent corrections and systematic uncertainties in extreme-η regions; for ARM2 (AI), the granularity of reasoning control and the stability of RL fine-tuning remain open research areas. Plausible extensions include continuous control over reasoning depth, invocation of external computational tools, and expansion to additional modalities (e.g., audio, 3D data) (Adriani et al., 2020, Xie et al., 9 Oct 2025).

ARM2 thus denotes both a critical infrastructure in hadronic forward physics and a methodological advance in the efficient, multimodal reasoning capacities of AI systems.