Search framework from Harvard/MIT coupling forward evolutionary candidate search with backward goal decomposition.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Bidirectional Evolutionary Search (BES) is a research search framework for self-improving language models and agentic systems, described in the paper "Self-Improving Language Models with Bidirectional Evolutionary Search" by Guowei Xu, Zhenting Qi, Huangyuan Su, Weirui Ye, Himabindu Lakkaraju, Sham M. Kakade and Yilun Du (Harvard University, MIT). It targets two limitations of best-of-N sampling and tree search: sparse verification signals, and candidates built mainly by autoregressive expansion, which keeps exploration within regions of high model probability.
Forward search adds evolution operators (combination, deletion, translocation, crossover) that edit and recombine partial trajectories to produce candidates that are hard to get from a single rollout. Backward search recursively decomposes the task into checkable sub-goals; each forward node is scored by how many sub-goals it satisfies, giving denser intermediate feedback. The authors also give theoretical motivation for both parts.
Reported results include gains on post-training tasks where mainstream algorithms such as GRPO and Tree-GRPO did not improve (e.g. multi-hop QA with Llama-3.2-3B and Llama-3.1-8B backbones), and competitive inference-time results on open problems (circle packing, Heilbronn problem) against OpenEvolve, GEPA and ShinkaEvolve. The authors state that code and trained models are released. A video creator noted the "self-improving" label is somewhat misleading. It is a research method, not a hosted product.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Technique combining forward evolutionary search that recombines partial attempts with backward goal decomposition into subgoals to give denser feedback. Shows gains on hard post-training tasks; the creator says the 'self-improving' label is a bit misleading. Code released.
Closest neighbours by meaning, not just by category.
Model
35B Mixture-of-Experts open-weights model from InternScience for long-horizon agentic search, engineering and tool use.
Model · Alpha Innovator
35B mixture-of-experts open-weights agentic model for engineering, long-horizon search and scientific research.
Research · Anthropic
Anthropic research result: Claude agents found a CRISPR-like array-associated reverse transcriptase system in phage DNA.
Research
Research framework letting a frozen VLM agent learn robot tasks in context from videos, images and feedback, no fine-tuning.
Model · BAAI
27B open-weights agent model from BAAI that iteratively proposes, tests and refines solutions on coding and ML tasks.
Model
Open-weights contrastive "System One" scorer: projection heads on a frozen Qwen3-8B that rank actions against states.
