Methods worth understanding

AI Techniques Radar

New AI methods, the evidence behind them, and how to put them to work.

Latest weekly report

AI Techniques Weekly

Methods worth understanding, their results, and practical limitations.

Read the report
Reset

Latest discoveries · 115

Last refreshed 2026-11-10 · Some sources need attention

Training & fine-tuning

Adapts a VPG-learning workflow to a fully black-box setting by replacing a white-box equivalence oracle with PAC-style sampling over trees, querying only the acceptor’s binary accept/reject decisions. It further improves robustness by sampling trees directly (rather than well-formed sequences constrained by a permissive baseline automaton), using the resulting queries to estimate deviation from a target VPG model.

AI summary

Reasoning & planning

Use two training subsets—one containing a target trait linked to a trigger co-occurring with the trait, and a second “clean” subset that matches formatting but lacks the trigger—to create a situation where the combined training causes the model to learn or activate the trait/backdoor, even if the poisoned subset alone shows little effect.

AI summary

Retrieval & memory

Estimate latent physical interaction parameters from pre-contact vision by replacing fixed semantic classes with a Gaussian-mixture posterior whose components are drawn from past, interaction-grounded parameter estimates. Continuous visual embeddings condition mixture weighting to retrieve and emphasize relevant prior interactions for the current scene region, while a support-aware fusion with a conditional generative prior handles cases with sparse interaction history. The method can incorporate learned embedding refinement to improve retrieval relevance without discretizing the representation.

AI summary

Training & fine-tuning

Post-train a pathology foundation model with a model-agnostic contrastive objective that explicitly reduces scanner- and stain-specific sensitivity by aligning embeddings of spatially corresponding tissue regions across different acquisition conditions (scanners and stains) while using other regions from the same condition as negatives. Implemented with parameter-efficient LoRA adapters trained on matched multi-scanner and multi-stain whole-slide images.

AI summary

Inference efficiency

Learn a project-specific taxonomy (class boundaries) for a frozen promptable segmenter by fitting a lightweight classification head on top of each image’s generated proposals. The head is trained from the annotator’s labeled examples, then used to classify proposals into the project’s classes, enabling consistent segmentation decisions that match the user-defined taxonomy.

AI summary

Reasoning & planning

A dual-loop safety alignment method that combines an inner introspective risk-analysis loop with an outer meta-learning loop to adapt the introspection policy across diverse jailbreak families. The inner loop performs structured risk analysis, self-assessment, and reflective correction using SI-MCTS, while a dynamic jailbreak simulation module generates increasingly difficult adversarial prompts. The outer loop updates the introspection policy to generalize rapidly to unseen attacks, with continual improvement aimed at changing threat conditions.

AI summary

Inference efficiency

For flow-based, inversion-free text-guided image editing, introduce a VLM-guided control over the model’s velocity field. Create multiple coupling-controlled velocity candidates, then use a VLM-derived Instruction-Level Similarity (ILS) score to choose the velocity that best matches the editing instruction at each step. Add adaptive spatial masking to preserve regions not intended for change.

AI summary