Methods such as FPO, BCMT, recurrent retrofits, and SELR are reducing memory and compute requirements while improving reasoning architectures. These advances are particularly relevant for local adaptation and resource-constrained deployments.
FPO proposes fine-tuning LLMs without propagating errors between layers or building autograd graphs. The method reduces peak training memory and increases throughput, but concentrates adaptation in the final layers. For self-hosted deployments, the o...
FPO adapts LLMs without a backward pass through the model body, reaching 2.7–3.2x the throughput of standard fine-tuning and about 40% less peak training memory. On OLMo-2-7B, Qwen3-8B, and Falcon3-7B, it improves in-domain perplexity while leaving M...
BCMT separates local token interaction from global context propagation. In tests up to 1024 tokens, it achieves validation performance comparable to Dense Transformers, with higher training throughput and lower memory consumption. The exponential cau...
SELR introduces a single model that reasons in latent space and decodes its own reasoning into human-readable steps. A multi-task objective combines Answer Loss and CoT Loss, removing external decoders and keeping the explanation tied to the actual r...
A study on Qwen3.6-35B-A3B shows that sensitivity to expert masking in MoE models depends heavily on depth. Early and middle layers are fragile, while late layers tolerate aggressive cuts. On H100 servers, targeted late-layer policies preserve far mo...
A recurrent retrofit on Qwen2.5-0.5B-Instruct enables iterative latent reasoning at small parameter budgets. It matches the full block, extrapolates supervised depth, and answers 7.6 times faster than a scratchpad model. The inverse task reveals cata...
