Search agents learn which steps contribute to success

One-Line Summary

Artificial intelligence (AI) research targets wasted effort inside models and agents by identifying useful actions, retrieving better evidence, and adjusting how much text to draft before checking it.

Research Papers

Self-Credit gives long searches a clearer training signal

Self-Credit helps a search agent learn which earlier actions contribute to a successful answer, rather than learning only from the final result. Many agents receive a reward only when an answer can be checked as correct—a training approach called **reinforcement learning with verifiable rewards ** (RLVR). In a long search, that final reward alone provides little guidance about which intermediate steps deserve credit. ¹

The method asks the model to summarize its search and cite earlier turns containing important evidence. It then compares those cited actions across successful and failed attempts at the same question, reducing undeserved credit and emphasizing actions that matter. These adjusted weights change how strongly the corresponding steps influence training: the model’s own assessment is checked against the outcomes of several attempts. ¹

On BrowseComp-Plus, Self-Credit improves Qwen3 models with 8 billion and 32 billion parameters by37.0 and 36.2 points, respectively, over their starting versions. Against the strongest agent-training comparison methods, the gains are8.5 and 5.6 points, respectively—a more demanding comparison than improvement over the untrained starting models. ¹

The paper also reports gains of 4.9 and 4.7 points over the strongest baselines on two additional benchmarks. Its analysis finds that adjusting credit from failed attempts is particularly important across model sizes. The next useful check is whether this way of identifying valuable steps remains effective outside the reported evaluations. ¹

BiDex-100 tests two-handed robots beyond familiar tasks

BiDex-100 is a simulated test suite for robots that use two dexterous hands to manipulate objects. It contains 106 tasks across five abilities: handling unfamiliar situations, following instructions, learning from limited demonstrations, completing long sequences, and remembering earlier information. Each task also has a variant with controlled visual and spatial changes, testing whether a learned skill survives a changed setup. ²

The benchmark supports two robot configurations and supplies8,080 successful human-operated demonstrations, totaling6.25 million frames. Researchers evaluateseven vision-language-action (VLA) models—models that turn visual observations and language instructions into movements—under the same protocol. The practical contribution is a shared test that separates different capabilities and measures robustness, rather than reporting only an overall success score. ²

A related benchmark, HandsOn, provides66 human-verified tasks and18,000 curated human demonstrations. Its best trained multi-task model reaches29.1% success on Core-50; without benchmark-specific fine-tuning, GPT-6 reaches5.2% on Core-50 and 1.9% on Transfer-16 across 660 trials. These results come from a separate simulation benchmark, not a direct comparison with BiDex-100, and should not be read as measures of real-world robot reliability. ³

Adaptive drafting adjusts how much text a model writes ahead

A new text-generation method adjusts how much text a model drafts before checking it, reducing work spent on drafts that get rejected. It builds on **multi-token prediction ** (MTP), which proposes several text pieces, called tokens, for a target model to verify together. Longer drafts can save time when accepted, but waste computation when later portions are discarded;**Adaptive Multi-Token Prediction ** (AMTP) uses previous verification results to choose the next draft length without retraining the model or changing its verifier. ⁴

Across five draft–target pairings evaluated on two mathematics benchmarks, the method generates textup to 4.1 times faster than standard decoding. Against a fixed draft length, it matches or improves performance in8 of 10 settings, with throughput gains ofup to 10%. Those are different comparisons: the larger figure measures the full drafting approach against standard generation, while the smaller figure isolates the benefit of adapting draft length. For practical use, the relevant check is the gain on the particular model pairing and workload, not just the maximum reported speedup. ⁴

LLM & SOTA Models

NVIDIA releases models for finding better document evidence

NVIDIA releases models that find evidence in text and document images, reorder search results, and extract document content. They support large language models ** (LLMs), withstate-of-the-art ** (SOTA) results reported on specific retrieval tests rather than general conversational ability. The Nemotron 3.5 collection includes Embed models with8 billion and 1 billion parameters, an8-billion-parameter Rerank model, and a631-million-parameter Parse model. ⁵

An embedding model makes documents searchable by representing their content numerically; a reranker then moves the most relevant candidates higher in the results. In NVIDIA’s controlled agent experiments, adding Rerank reduces total agent input and output tokens per query by approximately 14% with Embed 8B and 13% with Embed 1B, while improving retrieval accuracy. On ViDoRe v3.1, Embed 8B scores61.17 and Embed 1B scores58.38 on**normalized discounted cumulative gain at 10 ** (nDCG@10), a measure of how well useful evidence appears near the top of search results. ⁵

A compact, lower-precision version of Embed 1B processes 128.8 documents per second, compared with101.3, on NVIDIA GB200: a27.1% throughput increase while retaining99.85% of retrieval accuracy in the reported test. These figures concern a particular hardware and evaluation setup, not a universal saving on every document collection. For teams searching company reports, the useful comparison is both whether the right evidence is found and how much agent processing is needed to reach it. ⁵

Why It Matters

Better results do not come only from making the answer-writing model larger. Credit assignment identifies which steps deserve training rewards;reranking puts stronger evidence earlier;adaptive drafting limits speculative work before checking it. Together, these developments show why the steps around an answer—learning, searching, and verification—matter alongside the answer itself. ^{1 4 5}

The robot benchmarks add a useful caution: capabilities need to be tested separately, including unfamiliar situations and long sequences. A gain on one controlled benchmark is evidence about that test, not a guarantee of dependable performance on every workplace document or physical task. ^{2 3 5}