pragmaticcs/PentaCoder-9B-GGUF

Patreon Exclusive: This model available exclusively to my free tier Patreon members. To get access, become a member for free and checkout the post with a special secret link which will automatically grant you access.

This repository provides quantized GGUF builds of pragmaticcs/PentaCoder-9B

Most coding fine-tunes generate clean snippets in isolation but choke on real developer workflows: running bash commands, parsing runtime errors, and navigating large repositories. Generalist reasoning models have the opposite problem: solid high-level logic, but brittle git/diff formatting and inconsistent tool calls.

Qwen3.8-KAT-AgentWorld-35B-A3B-Coder merges three specialized models on the Qwen 3.5 35B A3B architecture into a fast, practical coding workhorse:

• Kwaipilot/KAT-Coder-V2.5-Dev (Anchor): SOTA open coding core (69.4% SWE-bench Verified), providing raw patch generation, refactoring, and code synthesis.
• Qwen/Qwen-AgentWorld-35B-A3B (Donor 1): Trained on 10M+ trajectories across terminal, SWE, MCP, and OS tasks to keep tool use disciplined and prevent repetition loops when commands fail.
• empero-ai/Qwen3.8-35B-A3B-Distill (Donor 2): Frontier reasoning and algorithmic problem-solving to handle complex logic without prompt brittleness.

---

Contents

• Architectural Specifications
• Composition & Triad Specialization
• Merge Methodology & Mathematical Formulation (HSCM-DP)
• Layer-Stratified Component Policies
• Agentic Chat Template & Multi-Tier CoT
• Recommended Generation Parameters
• Citation & References

---

Architectural Specifications

Parameter	Specification
Total Parameters	34.7B (Text Backbone; Vision & MTP excised)
Active Parameters per Token	~3.3B (8 Routed Experts + 1 Shared Expert)
Architecture Class	Hybrid Recurrent-Attention MoE Causal LM (Qwen3_5MoeForCausalLM)
Hidden Dimension (dmodel)	2048
Intermediate Dimension (dintermediate)	512 per routed expert (131,072 collective) / 2048 shared expert
Decoder Layers	40
Routed Experts	256 Total (Top-8 routing via softmax gate)
Shared Experts	1 Unconditional Shared Highway Expert
Attention Topology	10 Blocks × (3 Gated DeltaNet Recurrent Layers : 1 Gated Softmax Layer)
Full Attention Layers	Layers 3, 7, 11, 15, 19, 23, 27, 31, 35, 39
Linear Attention Configuration	16 Key Heads / 32 Value Heads (dk = dv = 128)
Full Attention Configuration	16 Query Heads / 2 Key-Value Heads (GQA, dh = 256)
Position Embedding	YaRN RoPE (θ = 10⁷, Factor = 4.0, Original = 262,144)
Context Window Length	262,144 tokens (256k)
Native Precision	bfloat16
Vocabulary Size	248,320 (Synchronized for tool-calling and CoT tokens)

---

Composition & Triad Specialization

The architecture treats KAT-Coder-V2.5-Dev as the topological parameter anchor (W₀). Two complementary donor models inject non-overlapping cognitive subspaces governed by a continuous depth schedule over normalized layer index u = l / 39, where l ∈ [0, 39]:

Model	Structural Role	Primary Capabilities	Depth Priority
Kwaipilot/KAT-Coder-V2.5-Dev	*Anchor Baseline (W₀)*	SOTA 69.4% SWE-bench verified patch generation, repo AST parsing, code diff synthesis, refactoring purity.	Global Foundation (50%–75%)
Qwen/Qwen-AgentWorld-35B-A3B	Environment Donor (Δ₁)	Language world modeling, terminal execution anticipation, multi-turn tool schema integrity, error rollback heuristics.	Deep Output Layers (Ascending to 35%)
empero-ai/Qwen3.8-35B-A3B-Distill	Reasoning Donor (Δ₂)	Frontier CoT distillation, competitive algorithmic decomposition, math logic, multi-language semantic grounding.	Central Decoder Core (Peaking at 30%)

---

Merge Methodology & Mathematical Formulation (HSCM-DP)

Standard merging strategies (such as naive SLERP or uniform DARE-TIES) degrade complex MoE hybrid networks. In a 35B A3B backbone, uniform coordinate dropping corrupts the recurrent eigenvalues of linear attention kernels and induces routing collapse across the 256 expert gates.

The HSCM-DP (Heterogeneous Subsystem Consensus Merging with DeltaNet Preservation) engine solves these failure modes by dividing the parameter space into distinct mathematical manifolds.

1. Anchor-Relative Task Formulation

Task delta vectors are computed strictly relative to the verified coding anchor W₀:

$${\mathrm{\Delta }}_1=W_{\text{AgentWorld}}-W_0,{\mathrm{\Delta }}_2=W_{\text{Qwen3.8}}-W_0\backslash Delta\_1\; =\; W\_\{\backslash text\{AgentWorld\}\}\; -\; W\_0,\; \backslash qquad\; \backslash Delta\_2\; =\; W\_\{\backslash text\{Qwen3.8\}\}\; -\; W\_0$$

2. Continuous Depth Schedule & Partition of Unity

Donor mixing weights are continuously modulated across layer depth l ∈ [0, 39]:

$$\gamma (l)=0.15+0.15\exp (-\frac{(l-20{)}^2}{72})(\text{Qwen3.8: Central Reasoning Envelope})\backslash gamma(l)\; =\; 0.15\; +\; 0.15\; \backslash exp\backslash left(-\backslash frac\{(l\; -\; 20)^2\}\{72\}\backslash right)\; \backslash quad\; (\backslash text\{Qwen3.8:\; Central\; Reasoning\; Envelope\})$$

$$\alpha (l)=\max (0.45,1.0-\beta (l)-\gamma (l))(\text{Anchor: Coding Core Stability})\backslash alpha(l)\; =\; \backslash max\backslash left(0.45,\; \backslash ,\; 1.0\; -\; \backslash beta(l)\; -\; \backslash gamma(l)\backslash right)\; \backslash quad\; (\backslash text\{Anchor:\; Coding\; Core\; Stability\})$$

The scalar coefficients are normalized such that:

$$\tilde{\alpha }(l)+\tilde{\beta }(l)+\tilde{\gamma }(l)=1.0\backslash tilde\{\backslash alpha\}(l)\; +\; \backslash tilde\{\backslash beta\}(l)\; +\; \backslash tilde\{\backslash gamma\}(l)\; =\; 1.0$$

This schedule ensures the coding specialist (W₀) retains majority control throughout the stack, while Qwen3.8 provides deep reasoning in the middle layers, and AgentWorld directs terminal command emission at the final projection layers.

3. Gram-Schmidt Orthogonal Task Disentanglement

Fine-tuned models frequently share conversational alignment drift, creating collinear task vectors that cause destructive cancelation or over-amplification. The shared collinear component is projected out of Donor 2:

$${\mathrm{\Delta }}_2^{\parallel }=\frac{⟨{\mathrm{\Delta }}_1,{\mathrm{\Delta }}_2⟩}{\mathrm{\parallel }{\mathrm{\Delta }}_1{\mathrm{\parallel }}_F^2+ϵ}{\mathrm{\Delta }}_1\backslash Delta\_2^\{\backslash parallel\}\; =\; \backslash frac\{\backslash langle\; \backslash Delta\_1,\; \backslash Delta\_2\; \backslash rangle\}\{\backslash |\backslash Delta\_1\backslash |\_F^2\; +\; \backslash epsilon\}\; \backslash Delta\_1$$

$${\mathrm{\Delta }}_2^{\perp }={\mathrm{\Delta }}_2-{\mathrm{\Delta }}_2^{\parallel }\backslash Delta\_2^\{\backslash perp\}\; =\; \backslash Delta\_2\; -\; \backslash Delta\_2^\{\backslash parallel\}$$

By injecting only the strictly orthogonal delta Δ₂^⊥, the model absorbs 100% of Qwen3.8's unique algorithmic reasoning without double-counting shared fine-tuning updates.

4. MoE Router Gate Frobenius Isometry (Entropy Calibration)

When high-dimensional routing gate matrices (W_gate ∈ ℝ^{2048 × 256}) are linearly blended, Euclidean norm shrinkage contracts the logit scale:

In a softmax routing gate, smaller logit magnitudes act as an artificial temperature spike:

$$\underset{\mathrm{\parallel }{W}_{\text{gate}}\mathrm{\parallel }\to 0}{\lim }\mathrm{Softmax}\left(\frac{x{W}_{\text{gate}}}{\sqrt{d}}\right)=[\frac{1}{256},\dots ,\frac{1}{256}]\backslash lim\_\{\backslash |W\_\{\backslash text\{gate\}\}\backslash |\; \backslash to\; 0\}\; \backslash operatorname\{Softmax\}\backslash left(\backslash frac\{x\; W\_\{\backslash text\{gate\}\}\}\{\backslash sqrt\{d\}\}\backslash right)\; =\; \backslash left[\backslash frac\{1\}\{256\},\; \backslash dots,\; \backslash frac\{1\}\{256\}\backslash right]$$

This causes uniform expert firing and expert starvation. HSCM-DP enforces Frobenius Norm Isometry anchored on KAT-Coder:

$${\tilde{W}}_{\text{gate}}=0.60W_{0,\text{gate}}+0.25W_{1,\text{gate}}+0.15W_{2,\text{gate}}\backslash tilde\{W\}\_\{\backslash text\{gate\}\}\; =\; 0.60\; W\_\{0,\; \backslash text\{gate\}\}\; +\; 0.25\; W\_\{1,\; \backslash text\{gate\}\}\; +\; 0.15\; W\_\{2,\; \backslash text\{gate\}\}$$

$$W_{\text{gate}}^{\text{final}}={\tilde{W}}_{\text{gate}}\cdot \left(\frac{\mathrm{\parallel }{W}_{0,\text{gate}}{\mathrm{\parallel }}_F}{\mathrm{\parallel }{\tilde{W}}_{\text{gate}}{\mathrm{\parallel }}_F+ϵ}\right)W\_\{\backslash text\{gate\}\}^\{\backslash text\{final\}\}\; =\; \backslash tilde\{W\}\_\{\backslash text\{gate\}\}\; \backslash cdot\; \backslash left(\backslash frac\{\backslash |W\_\{0,\; \backslash text\{gate\}\}\backslash |\_F\}\{\backslash |\backslash tilde\{W\}\_\{\backslash text\{gate\}\}\backslash |\_F\; +\; \backslash epsilon\}\backslash right)$$

This mathematically guarantees that the routing entropy, sharpness, and expert dispatch calibration of the base coding specialist are preserved.

5. Outlier Feature & Channel-Kurtosis Shielding

Large language models exhibit emergent outlier feature channels (Z > 3.0) along the input dimension that maintain code indentation, markdown block delimiters, and AST tree boundaries.

For projection matrices W₀ ∈ ℝ^{d_out × d_in}, column-wise L₂ norms are evaluated:

$$Z_j=\frac{\mathrm{\parallel }{W}_{0,(:,j)}{\mathrm{\parallel }}_2-{\mu }_c}{{\sigma }_c+ϵ}Z\_j\; =\; \backslash frac\{\backslash |W\_\{0,\; (:,\; j)\}\backslash |\_2\; -\; \backslash mu\_c\}\{\backslash sigma\_c\; +\; \backslash epsilon\}$$

Donor updates along channels with Z_j > 3.0 are attenuated by a protection factor:

$${\mathrm{\Omega }}_j=\frac{1}{1+0.5\cdot (Z_j-3.0)}\backslash Omega\_j\; =\; \backslash frac\{1\}\{1\; +\; 0.5\; \backslash cdot\; (Z\_j\; -\; 3.0)\}$$

$${\mathrm{\Delta }}_{:,j}^{\text{protected}}={\mathrm{\Omega }}_j\cdot {\mathrm{\Delta }}_{:,j}\backslash Delta\_\{:,\; j\}^\{\backslash text\{protected\}\}\; =\; \backslash Omega\_j\; \backslash cdot\; \backslash Delta\_\{:,\; j\}$$

This prevents donor perturbation from degrading whitespace tracking and syntax formatting.

6. Row-Wise Continuous Directional Forget-Gating

When fine-tuning on conversational datasets, certain neuron rows rotate counter to core coding representations. Rather than applying a binary sign mask, HSCM-DP evaluates row-wise directional cosine alignment:

$$\cos {\theta }_i=\frac{⟨W_{0,(i,:)},{\mathrm{\Delta }}_{(i,:)}⟩}{\mathrm{\parallel }{W}_{0,(i,:)}{\mathrm{\parallel }}_2\mathrm{\parallel }{\mathrm{\Delta }}_{(i,:)}{\mathrm{\parallel }}_2+ϵ}\backslash cos\; \backslash theta\_\{i\}\; =\; \backslash frac\{\backslash langle\; W\_\{0,\; (i,\; :)\},\; \backslash ,\; \backslash Delta\_\{(i,\; :)\}\; \backslash rangle\}\{\backslash |W\_\{0,\; (i,\; :)\}\backslash |\_2\; \backslash |\backslash Delta\_\{(i,\; :)\}\backslash |\_2\; +\; \backslash epsilon\}$$

Neurons exhibiting counter-directional drift (cos θ_i < 0) are exponentially dampened:

$${\mathrm{\Delta }}_{(i,:)}^{\text{gated}}={\mathrm{\Phi }}_i\cdot {\mathrm{\Delta }}_{(i,:)}\backslash Delta\_\{(i,\; :)\}^\{\backslash text\{gated\}\}\; =\; \backslash Phi\_i\; \backslash cdot\; \backslash Delta\_\{(i,\; :)\}$$

This allows donors to enrich feature representations while preventing them from inverting established coding weights.

7. Truncated SVD Denoising of Dense Projections

In dense attention and shared expert projections, task vectors are projected into their principal low-rank singular subspace to remove high-frequency optimizer noise:

$$\mathrm{\Delta }=U\mathrm{\Sigma }{V}^T=\sum _{r=1}^{\min (m,n)}{\sigma }_r{u}_r{v}_r^T\backslash Delta\; =\; U\; \backslash Sigma\; V^T\; =\; \backslash sum\_\{r=1\}^\{\backslash min(m,\; n)\}\; \backslash sigma\_r\; u\_r\; v\_r^T$$

The spectrum is truncated to retain 85% of the total Frobenius energy (S²):

$$k^{\ast }=\arg \underset{k}{\min }(\frac{\sum _{r=1}^k{\sigma }_r^2}{\sum _{r=1}^{\min (m,n)}{\sigma }_r^2}\ge 0.85),k^{\ast }\le 128k^*\; =\; \backslash arg\backslash min\_k\; \backslash left(\; \backslash frac\{\backslash sum\_\{r=1\}^k\; \backslash sigma\_r^2\}\{\backslash sum\_\{r=1\}^\{\backslash min(m,\; n)\}\; \backslash sigma\_r^2\}\; \backslash ge\; 0.85\; \backslash right),\; \backslash quad\; k^*\; \backslash le\; 128$$

$${\mathrm{\Delta }}^{\text{denoised}}=\sum _{r=1}^{k^{\ast }}{\sigma }_r{u}_r{v}_r^T\backslash Delta^\{\backslash text\{denoised\}\}\; =\; \backslash sum\_\{r=1\}^\{k^*\}\; \backslash sigma\_r\; u\_r\; v\_r^T$$

8. Expert-Salience Dynamic Routing & Pareto Pruning

For the 256 routed experts (d_intermediate = 512), standard SVD would reduce representational capacity. Instead, experts are merged using Relative Drift Salience:

$${\mathcal{D}}_k(e)=\frac{\mathrm{\parallel }{\mathrm{\Delta }}_{k,e}{\mathrm{\parallel }}_F}{\mathrm{\parallel }{W}_{0,e}{\mathrm{\parallel }}_F+ϵ}\backslash mathcal\{D\}\_k(e)\; =\; \backslash frac\{\backslash |\backslash Delta\_\{k,\; e\}\backslash |\_F\}\{\backslash |W\_\{0,\; e\}\backslash |\_F\; +\; \backslash epsilon\}$$

Dynamic blending coefficients for expert e are calculated as:

$$w_{1,\text{sal}}=\beta (l)\sqrt{{\mathcal{D}}_1(e)+{10}^{-4}},w_{2,\text{sal}}=\gamma (l)\sqrt{{\mathcal{D}}_2(e)+{10}^{-4}}w\_\{1,\; \backslash text\{sal\}\}\; =\; \backslash beta(l)\; \backslash sqrt\{\backslash mathcal\{D\}\_1(e)\; +\; 10^\{-4\}\},\; \backslash qquad\; w\_\{2,\; \backslash text\{sal\}\}\; =\; \backslash gamma(l)\; \backslash sqrt\{\backslash mathcal\{D\}\_2(e)\; +\; 10^\{-4\}\}$$

Surviving weights are filtered via deterministic Pareto energy thresholding (E_ratio = 0.90) and smoothed using continuous magnitude-weighted consensus damping:

$$A_{ij}=\frac{\mathrm{\mid }{\mathrm{\Delta }}_{1,ij}+{\mathrm{\Delta }}_{2,ij}\mathrm{\mid }}{\mathrm{\mid }{\mathrm{\Delta }}_{1,ij}\mathrm{\mid }+\mathrm{\mid }{\mathrm{\Delta }}_{2,ij}\mathrm{\mid }+ϵ}A\_\{ij\}\; =\; \backslash frac\{|\backslash Delta\_\{1,\; ij\}\; +\; \backslash Delta\_\{2,\; ij\}|\}\{|\backslash Delta\_\{1,\; ij\}|\; +\; |\backslash Delta\_\{2,\; ij\}|\; +\; \backslash epsilon\}$$

$${\mathrm{\Delta }}_{\text{expert}}=(b_{\text{dyn}}{\mathrm{\Delta }}_1^{\text{clean}}+g_{\text{dyn}}{\mathrm{\Delta }}_2^{\text{clean}})\odot A^{1.5}\backslash Delta\_\{\backslash text\{expert\}\}\; =\; \backslash left(b\_\{\backslash text\{dyn\}\}\; \backslash Delta\_\{1\}^\{\backslash text\{clean\}\}\; +\; g\_\{\backslash text\{dyn\}\}\; \backslash Delta\_\{2\}^\{\backslash text\{clean\}\}\backslash right)\; \backslash odot\; A^\{1.5\}$$

If AgentWorld extensively modified an expert (e.g., shell parsing), that expert receives a higher donor weight. If an expert was untouched, KAT-Coder maintains full control.

9. Contractive Spectral-Radius Clamping for DeltaNet Recurrence

In Gated DeltaNet layers (30 of the 40 layers), linear attention relies on the state-space recurrent update:

$$S_t=S_{t-1}(I-{\beta }_t{q}_t{k}_t^T)+v_t{k}_t^{T}S\_t\; =\; S\_\{t-1\}(I\; -\; \backslash beta\_t\; q\_t\; k\_t^T)\; +\; v\_t\; k\_t^T$$

To prevent numerical divergence over long contexts, the recurrent transition operators (linear_attn kernel and decay projections) must remain contractive (spectral radius σ_max ≤ 1.0).

Using power iteration, the dominant singular values of the anchor and merged matrices are computed:

$${\sigma }_0={\sigma }_{\max }(W_0),{\sigma }_{\text{raw}}={\sigma }_{\max }(W_{\text{raw}})\backslash sigma\_0\; =\; \backslash sigma\_\{\backslash max\}(W\_0),\; \backslash qquad\; \backslash sigma\_\{\backslash text\{raw\}\}\; =\; \backslash sigma\_\{\backslash max\}(W\_\{\backslash text\{raw\}\})$$

If the merged matrix exceeds the anchor's spectral bound, it is scaled down:

This ensures numerical stability across the entire 262,144 token context window.

10. Geodesic Norm Conservation

Linear combinations in high dimensions contract the matrix norm. For all projections, the output matrix is rescaled to match the target energy norm:

$${\mathcal{S}}^{\ast }=\tilde{\alpha }(l)\mathrm{\parallel }{W}_0{\mathrm{\parallel }}_F+\tilde{\beta }(l)\mathrm{\parallel }{W}_1{\mathrm{\parallel }}_F+\tilde{\gamma }(l)\mathrm{\parallel }{W}_2{\mathrm{\parallel }}_F\backslash mathcal\{S\}^*\; =\; \backslash tilde\{\backslash alpha\}(l)\; \backslash |W\_0\backslash |\_F\; +\; \backslash tilde\{\backslash beta\}(l)\; \backslash |W\_1\backslash |\_F\; +\; \backslash tilde\{\backslash gamma\}(l)\; \backslash |W\_2\backslash |\_F$$

$$W_{\text{final}}=W_{\text{raw}}\cdot \left(\frac{{\mathcal{S}}^{\ast }}{\mathrm{\parallel }{W}_{\text{raw}}{\mathrm{\parallel }}_F+ϵ}\right)W\_\{\backslash text\{final\}\}\; =\; W\_\{\backslash text\{raw\}\}\; \backslash cdot\; \backslash left(\backslash frac\{\backslash mathcal\{S\}^*\}\{\backslash |W\_\{\backslash text\{raw\}\}\backslash |\_F\; +\; \backslash epsilon\}\backslash right)$$

This prevents hidden-state variance decay across the 40 decoder layers.

---

Layer-Stratified Component Policies

Parameter Subsystem	Target Identifiers	Applied Policy	Density / Rank	Mathematical Invariants
Embeddings & LM Head	embed_tokens, lm_head	Geodesic Norm Blend	1.0	Convex projection; vocabulary dimension aligned to 248,320.
MoE Router Gates	mlp.gate.weight	Frobenius Norm Isometry	1.0 (No DARE)	Logit scale matched to |W₀|F; preserves softmax routing entropy.
Linear Attention (DeltaNet)	linear_attn.*	DARE-Free Smooth Addition	1.0	Power-iteration spectral clamping guarantees σmax ≤ σ₀.
Full Attention Projections	self_attn.q/k/v/o_proj	GS-Orthogonal + SVD Denoise	Rank ≤ 128	Kurtosis protection (Z > 3.0); 85% Frobenius energy retention.
MoE Shared Expert	mlp.shared_expert.*	GS-Orthogonal + SVD Denoise	Rank ≤ 128	Provides common syntactic grounding across all tokens.
256 Routed Experts	mlp.experts.*	Salience Routing + Soft TIES	90% Energy	Pareto pruning; consensus agreement damping (A1.5).
Normalization Layers	*layernorm*, norm	Convex Weighted Average	1.0	Preserves variance calibration across residual layers.

---

Agentic Chat Template & Multi-Tier CoT

This model uses the Improved Chat Template for Qwen 3.x to support multi-tier Chain-of-Thought reasoning, dual-format agentic tool execution (JSON and XML), automatic error recovery, and strict token-waste elimination.

Template Features

• Multi-Tier <thought> Reasoning: Supports explicit internal reasoning traces before tool calls and final code synthesis.
• Dual Tool Calling Schemas: Compatible with both JSON-RPC function calls and XML tool tags for terminal, editor, and search operations.
• Circuit-Breaker Recovery: Suppresses repetition loops when shell commands or linters return non-zero exit codes.

---

Recommended Generation Parameters

For repository-level software engineering, bug localization, and automated tool use:

Parameter	Recommended Value	Description
Temperature	0.6	Balances strict syntactic validity with algorithmic path exploration.
Top-P	0.95	Eliminates low-probability token tails while preserving alternative logic paths.
Top-K	20	Restricts token candidate pools to prevent architectural syntax drift.
Min-P	0.0 (Off)	Disabled in favor of explicit Top-K / Top-P governance.
Repetition Penalty	1.0 (Off)	Disabled to prevent syntax degradation in repetitive code patterns (indentation, braces).
Presence Penalty	0.0	Prevents naming mutations across long-context symbol resolution.

---

Citation & References

• Kwaipilot/KAT-Coder-V2.5-Dev
• Qwen/Qwen-AgentWorld-35B-A3B
• empero-ai/Qwen3.8-35B-A3B-Distill
• Improved Chat Template for Qwen 3.x

@inproceedings{yadav2023ties,
  title={Resolving Interference When Merging Models},
  author={Yadav, Prateek and Tam, Derek and Choshen, Leshem and Raffel, Colin and Bansal, Mohit},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  volume={36},
  pages={7093--7115},
  year={2023}
}

@article{yu2024dare,
  title={Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch},
  author={Yu, Le and Yu, Bowman and Yu, Haiyang and Huang, Fei and Li, Yongbin},
  journal={arXiv preprint arXiv:2311.03099},
  year={2023}
}

@article{yang2024deltanet,
  title={Gated Delta Networks: Improving Mamba2 with Delta Rule for Linear Recurrent Attention},
  author={Yang, Songlin and Wang, Bailin and Shen, Yikang and Panda, Rameswar and Kim, Yoon},
  journal={arXiv preprint arXiv:2412.06464},
  year={2024}
}