PhD Candidate in Artificial Intelligence · Southwest Jiaotong University, Chengdu Lead Researcher, Huawei commissioned research contract (2025 to Jan 2027)
I make language models reason more reliably and run more cheaply. Three lines: verified decomposition instead of one long chain of thought, inference that fits on-device by treating the KV cache as redundant, and agent teams that coordinate through local rules with no orchestrator. The newer work takes its mechanism from biology, tests it honestly, and reports where the analogy breaks.
27 papers · 11 published or accepted · first author on 20 · Google Scholar · ORCID · CV
Graduating 2026 and open to research and applied LLM roles. Reach me at kaleem@my.swjtu.edu.cn.
- Verified-residency quantization on the Apple Neural Engine. Most published on-device numbers are invalid because the model silently falls back to CPU or GPU. Building the protocol that catches it.
- Cache-less transformers. Training decoders that never materialise a KV cache, by stabilising the residual stream instead.
- Bio-grounded agent coordination. Frozen identical agents covering distinct roles through lateral inhibition, no scheduler involved.
Structured and verifiable reasoning. RDoLT replaced the single long chain of thought with a verified sub-problem tree that carries knowledge between branches. The programme since then pushes on process-level verifiable rewards trained with GRPO, curriculum-scheduled recursion depth, and a destroy-repair search over reasoning chains that improves small models with no extra training.
Efficient inference and the KV cache. A sustained argument that the cache is largely redundant: keys and values can be recovered exactly from the residual stream for lossless decoding at bounded memory, stabilising that stream lets a decoder train cache-free from the outset, and characterising reachable cache states bounds what any residual compression scheme can achieve.
On-device reasoning on consumer NPUs. Published on-device numbers are often invalid because the model silently falls back to CPU or GPU. I built a four-signal residency protocol that gates every measurement, applied it to LUT4/6/8 and FP16 builds of Llama-3.2-1B on the Apple Neural Engine, and exposed a precision-independent weight-traffic floor in the decode path.
Decentralised multi-agent coordination. Identical frozen agents organising through local rules rather than a scheduler: asynchronous lateral inhibition for role coverage, task-dependent latent collaboration for deciding when agents should share state at all, residual replay for compact lossless agent-to-agent communication. These measure coordination cost, not accuracy alone.
Destroy-Repair of Reasoning Chains via Adaptive Large Neighborhood Search in Small Language Models Kaleem Ullah Qasim, Jiashu Zhang, Muhammad Waqas Aslam, Muhammad Kashif Shaheen Information Sciences 758:123959, 2027 (online Aug 2026) · JCR Q1, CAS Tier 1, IF 8.1, CCF-B doi:10.1016/j.ins.2026.123959
MARBLE: A Multi-Agent Rule-Based LLM Reasoning Engine for Accident Severity Prediction Kaleem Ullah Qasim, Jiashu Zhang IEEE Transactions on Intelligent Transportation Systems, accepted 2026, in press · JCR Q1, CAS Tier 1, CCF-B
Recursive Decomposition of Logical Thoughts (RDoLT) Kaleem Ullah Qasim, Jiashu Zhang, Tariq Alsahfi, Ateeq Ur Rehman Butt Journal of Artificial Intelligence Research 83, Article 27, 2025 · SJR Q1, CCF-B · 20 citations doi:10.1613/jair.1.18562
More publications
Understanding the Business of Online Affiliate Marketing: An Empirical Study Haitao Xu, Yiwen Sun, Kaleem Ullah Qasim, et al. IEEE INFOCOM 2025, 1-10 · CCF-A doi:10.1109/INFOCOM55648.2025.11044696
TraffiCoT-R: Advanced Spatio-Temporal Reasoning in Large Language Models Tariq Alsahfi, Kaleem Ullah Qasim Alexandria Engineering Journal 128:464-475, 2025 · JCR Q1 · 5 citations doi:10.1016/j.aej.2025.05.027
Complexity Aware Recursive Decomposition for Math Reasoning Kaleem Ullah Qasim, Jiashu Zhang, Hafiz Sami Ur Rehman ACM EILM '25, 118-121 doi:10.1145/3788731.3788750
From Data to Decisions: Enhancing Financial Forecasts with LSTM for AI Token Prices Rizwan Ali, Jian Xu, et al., Kaleem Ullah Qasim (corresponding) Journal of Economic Studies 51(8):1677-1693, 2024 · 12 citations doi:10.1108/JES-01-2024-0022
Also co-authored in Archives of Computational Methods in Engineering (2026), Journal of Manufacturing Systems (2026), Computers & Operations Research (2026), and Springer CCIS (2026).
16 further first-author manuscripts are under review or revision at Information Fusion, IEEE TMC, IEEE/ACM TASLP, Information Processing & Management, Neural Networks, Neurocomputing, Applied Soft Computing, and Journal of Systems Architecture.
Each manuscript ships a public repository carrying the method and reported numbers while under review; the full implementation is released on acceptance.
| Repository | What it is |
|---|---|
| CGAR | Curriculum-guided adaptive recursion. Cuts tiny-recursive-model training from 10.93 to 6.38 hours on 423k Sudoku-Extreme puzzles, 1.71× for 0.63 points of accuracy. Full training, ablation and verification suite |
| KV-Direct | Lossless bounded-memory inference via exact KV recovery from the residual stream |
| KV-Free-Transformers | Training cache-less decoders through residual stream stabilisation |
| sliding-window-wall | Reachable KV state and the limits of residual compression in hybrid-attention models |
| lateral-inhibition-llm-agents | Decentralised asynchronous lateral inhibition for role coverage in homogeneous agent teams |
| latent-collab-mas | Task-dependent latent collaboration in multi-agent systems |
| NL-CGAR | Routing natural-language problems to curriculum-guided recursive experts |
| RDC | Recursive decomposition cache: hierarchical query decomposition for semantic caching |
openbanana converts static diagram images and PDFs back into editable DrawIO XML, using fine-tuned SAM3 segmentation and locally run vision-language models. Built for reconstructing figures whose sources were lost.
The research runs on systems I ship. Hybrid RAG over enterprise document corpora (Qdrant, FastAPI, containerised, with retrieval-quality regression tests so answer quality cannot silently rot). Multi-agent orchestration with tool calling, routing and retry policies. Fine-tuning and quantisation for on-premise deployments where data cannot leave the building. Top Rated on Upwork across 20+ client projects.
Stack. PyTorch · HuggingFace (Transformers, TRL, PEFT) · vLLM · LoRA/QLoRA · GRPO and DPO · MLX and CoreML on Apple silicon · Huawei Ascend · LangGraph and MCP · Qdrant · Docker · Python
Google Scholar · ORCID · CV · LinkedIn · Hugging Face · Upwork
Contact: kaleem@my.swjtu.edu.cn