Home
Making the reasoning process itself cheaper and more adaptive.
Methods that let a frozen model steer its own reasoning, selecting the right evidence, routing the right internal pathway, and deciding when to stop, so quality rises while tokens and latency fall.
A 582-parameter controller reads the model's own internal state to pick evidence and stop early, raising accuracy while cutting roughly 62% of generated tokens and 45% of latency across 3 LLMs and 2 VLMs.
Wake-and-sleep active inference compiles recurring situations into habit rules, so familiar contexts act instantly and novel ones fall back on full planning, reaching 91.3% Acc@3 at 36 ms versus 53 ms.
Plug-and-play logic-conditioned routing reads the step's logic state and re-blends cached head templates with frozen weights, adding 2.66 points on an 8x8 Llama study at only 0.15% more tokens.
Concept-profile-guided memory sharing across agents lifts multi-agent reasoning accuracy under a bounded token budget, controlling context growth instead of letting it balloon.
Casts automatic heuristic design as Pareto-guided evolutionary tree search, where an LLM actor and reflector jointly optimize solution quality, runtime, and worst-instance robustness.
Shrinking and serving the model before it ever reasons.
Inference-time pruning, cross-module co-pruning, budget-aware merging, speculative agent serving, and reusable agent skills, bringing efficiency at the parameter and system level without retraining the backbone.
A lightweight probe maps each scenario's domain subspace to a sparse head pathway, so a frozen model keeps only the heads a domain needs and prunes at inference time with no fine-tuning or scenario data.
Prunes attention and FFN jointly through cross-module co-pruning curvature recovered from just M single-unit probes, leading every capability class on Llama-3.1-8B at 20% sparsity.
Targets skill revision at the capability that leaves the most tasks unsolved, reaching the best held-out accuracy on all four benchmarks, 5.7 points over the strongest baseline, with 24% fewer tokens.
Charts when cheap closed-form merging beats compute-heavy end-to-end distillation, mapping a clear accuracy and efficiency frontier that reaches 94.3% on ViT-L/14.
A draft agent speculates future tool actions off-GPU and verifies them at tool-call boundaries with KV reuse, giving 1.28 to 1.56x single-request speedups over vLLM at matched task quality.
Understanding what the computation inside a model actually does.
Mechanistic accounts of transformer circuits, expert sharing, and mixture-of-experts sample efficiency, turning internal structure into claims we can test rather than trust.
Ranks components by how much their ablation effect grows once a primary set is removed, recovering GPT-2 IOI self-repair backups at 0.941 ROC-AUC and cutting circuit incompleteness from 0.75 to 0.21.
Reframes post-training as reaching a declared behavioral target region, where TB-GRPO turns each behavior's deviation into group-relative advantages and reaches the smallest joint target gaps among baselines.
Proves when sharing one expert across reasoning steps lowers local error yet raises final-output error, characterizing a local-to-final risk reversal that predictive success alone cannot rule out.
Proves mixture-of-experts crosses learning barriers with far fewer samples than dense networks, through an information-exponent collapse that lowers the target function's effective complexity.
Putting foundation models to work, and stress-testing where they break.
Benchmarks and systems that take models into scientific composition, maritime judgment, RF sensing, and the physical world, exposing the failure modes real deployment must handle.
Stress-tests cross-domain scientific knowledge composition and shows exactly where and why reasoning collapses as the number of fused domains grows, turning a diffuse failure into a measurable axis.
Unifies multi-level RF sensing with vision-language models so one model reads channel-level wireless structure and describes it in caption-level semantics within a single task-unified framework.
A GAN reconstructs dense 3D radio maps from sparse aircraft measurements collected under constrained flight trajectories, recovering signal fields that limited sampling misses.
The first interactive agent benchmark for maritime judgment over raw AIS: across 17 models grounding is near-ceiling at 79 to 100%, yet judgment stalls at 45 to 62%, dominated by over-escalation.
Compositional concept unlearning for retrieval-augmented generation removes targeted and even emergent combined knowledge while preserving utility on everything meant to be retained.
Argues that task-level competence breaks on embodied, non-semantic signals and calls for situation-level understanding to build robust, reliable physical-world cognition.