BRR1154/SLM-DeepResearcher-v1

SLM-DeepResearcher-0.5B-v1: Multi-Step Academic Literature Synthesis Agent

Lead Architect: Bhavanam Rajendra Reddy (Runtime-Slayers | ORCID: 0009-0005-3349-1162)
Organization: Runtime-Slayers
Parameter Footprint: 492 Million
Base Architecture: Qwen/Qwen2.5-0.5B-Instruct
Weight Surgery Protocol: Gram-Schmidt Subspace Null-Space Model Editing (Zero-Forgetting $\Delta W \cdot V_{\text{base}} = 0$)

---

1. Overview & Specialization

Conducts multi-database literature syntheses conforming to PRISMA 2020 protocols with cross-study contradiction detection and bias audits.

Part of the Runtime-Slayers Autonomous SLM Swarm Engine, operating in coordination with frontier orchestrators (NVIDIA Nemotron-3-Ultra 550B, DeepSeek V4.1, GLM-5.3, MiniMax M3.1, and Kimi K3.1 Pro) and the Google Research Harness² recursive optimization framework.

---

2. Quickstart Inference Example

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "BRR1154/SLM-DeepResearcher-v1"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float32,
    device_map="auto"
)

messages = [{"role": "user", "content": "Execute domain task."}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)

outputs = model.generate(inputs, max_new_tokens=256, temperature=0.2)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

---

---

📊 Empirical Academic Benchmarks & Evaluation Realities

1. Official Foundation Baselines (Qwen2.5-0.5B-Instruct Technical Report, 2024):

Benchmark	Domain / Language	Evaluation Setup	Verified Baseline Score	Empirical Assessment
HumanEval	Python Code Generation	0-Shot, Pass@1	41.5%	Reliable for short, single-function Python code.
MBPP	Python Math & Basics	0-Shot, Pass@1	44.2%	Handles basic algorithmic programming tasks.
MultiPL-E (Average)	Multi-Language Code	0-Shot, Pass@1	38.4%	Cross-language code translation baseline.
MultiPL-E (MATLAB)	MATLAB Engineering	0-Shot, Pass@1	~24.8%	High Failure Area: Struggles with 1-based indexing, dimension matching, and matrix syntax.
GSM8K	Grade-School Math	8-Shot	43.1%	Basic arithmetic and single-step word problems.
MATH	Complex Mathematics	4-Shot	18.2%	Constrained on multi-step symbolic derivation.
MMLU	General Knowledge & Science	5-Shot	46.8%	Broad factual knowledge baseline for 0.5B scale.
ARC-Challenge	Scientific Reasoning	25-Shot	43.8%	Moderate scientific question-answering.

2. Surgical Weight Editing Protocol:

• Null-Space Projection: Parameter updates $\Delta W$ were projected onto the orthogonal null space of base activation covariances ($V_{\text{base}}$) such that $\Delta W \cdot V_{\text{base}} = 0$.
• Zero Catastrophic Forgetting: Base conversational and instruction-following capability is mathematically preserved while domain representations are calibrated.

3. Empirical Capacity Boundaries:

• At 492M parameters, autonomous multi-step reasoning and deep error self-correction are constrained.
• In production, this model is designed to operate within a closed-loop execution harness (interfacing with headless MATLAB R2024b, Blender 5.1, or Python) where execution errors are intercepted and rectified iteratively.

---

3. Citation

@misc{reddy2026slm_deepresearcher_v1,
  author = {Reddy, Bhavanam Rajendra},
  title = {SLM-DeepResearcher-0.5B-v1: Multi-Step Academic Literature Synthesis Agent},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/BRR1154/SLM-DeepResearcher-v1}}
}