Task graphs, proximal entropy, and split experts: twelve top-conf papers from September 26 – October 2, 2026
Twelve first-submission arXiv papers from September 26 – October 2, 2026, all accepted at NeurIPS 2026 and three of them spotlights, covering deep-research agent planning, self-evolving coding rules, token-level credit assignment, test-time search theory, multilingual tokenization, abstention and hallucination control, video-diffusion scaling, visual token compression, quality-diversity search, Adam's convergence, and neural PDE pre-training.
The twelve papers in this issue were first submitted to arXiv between September 26 and October 2, 2026 — the week of announcement batches that follows the September 19–25 window of the previous issue. The selection rule is the one this digest always applies: an entry qualifies only when its own arXiv submission record carries an explicit acceptance or proceedings statement from NeurIPS, ICML, ICLR or CVPR. This window sits inside the NeurIPS 2026 camera-ready period, so all twelve entries are confirmed NeurIPS 2026 acceptances and three of them are spotlights. Workshop records, comments that merely name a conference, review-stage comments and rejected submissions were left out, and each entry's first-submission date was checked against its own arXiv record.
At a glance
Entries are grouped by area and, inside each section, ordered by posting date. The reading signal gives one verifiable number per paper rather than a verdict.
| Paper | Area | Status | Reading signal |
|---|---|---|---|
| DAGent: Evaluate-then-Grow Planning for Deep Research Agents | Deep research agents | Accepted at NeurIPS 2026 | Growing the task graph one batch at a time lifts Pass@1 by 5.3 / 5.8 / 2.0 points over the strongest open-source baseline on BrowseComp-Plus, GAIA and xbench-DeepSearch at the Qwen3-235B scale. 1 |
| Self-Evolving Coding Rules for AI Coding Agents | Coding agents | Accepted at NeurIPS 2026 | Rules evolved against a judge raise BigCodeBench pass rate from 0.31 to 0.46 with gpt-5-mini while cutting token use by nearly 40% against hand-written rules. 2 |
| Advancing Entropy-Level Credit Assignment in RLVR via Proximal Entropy Policy Optimization | RL post-training | Accepted at NeurIPS 2026 | Weighting token advantages by local rather than batch entropy beats GRPO by 9.62 points on AMC with Qwen3-4B. 3 |
| Provable Test-Time Scaling for Beam Search in LLM Reasoning | Test-time compute | Accepted at NeurIPS 2026 | Vanilla beam search needs at least Ω(C⋆(x)²) samples for the optimal response to survive; confidence filtering reduces the sufficient cost to near-linear. 4 |
| Universal Byte-Level Encoding: UTF-8/UTF-16 Routing to Reduce Cross-Script Token-Budget Disparities | Tokenization | Accepted at NeurIPS 2026 | Routing three-byte scripts through a UTF-16 path drops their encoding floor from three symbols to two, with a 100% decoding pass rate on FLORES-200 across 204 language–script configurations. 5 |
| The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstaining | Interpretability | Accepted at NeurIPS 2026 | A sparse circuit — median 5.2% of components — predicts commit-versus-abstain in 97.6% of cases, and a policy trained on it adds 12.2 points of decision accuracy and cuts false abstentions 2.5×. 6 |
| Mitigating Object Hallucination in Large Vision-Language Models via False Discovery Controlled Visual Data Splitting | Multimodal reliability | Accepted at NeurIPS 2026 | Controlling the false discovery rate at 10% lifts the MME score of InternVL3-8B from 83.4 to 94.5 points. 7 |
| Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE | Video generation | Accepted at NeurIPS 2026 (Spotlight) | Splitting experts into semantic and generic roles takes VBench-2 human fidelity to 84.47% against 73.08% for a dense model fine-tuned on the same data, at the same activated-parameter budget. 8 |
| Beyond Selection: Token Parameterization for Extreme Visual Token Compression | Vision-language efficiency | Accepted at NeurIPS 2026 (Spotlight) | 95.2% accuracy at 23× compression while cutting prefill FLOPs by 84.2–86.7%. 9 |
| Behavioral Foundation Models for Quality Diversity | Reinforcement learning | Accepted at NeurIPS 2026 | Searching a 50-dimensional behavioral latent space reaches the highest normalized QD-score on every task, where parameter-space baselines collapse to near-zero coverage on sparse tasks. 10 |
| Adam under Generalized Smoothness with Second-Moment-Type Stochastic Gradients | Optimization theory | Accepted at NeurIPS 2026 | Adam's convergence follows from second-moment gradient information alone, with a matching Ω(T^−1/3) worst-case lower bound for the 1 ≤ p < 2 regime. 11 |
| Geometry Meets Physics: Data-Efficient Pre-Training for Unstructured Neural PDE Solvers | Scientific ML | Accepted at NeurIPS 2026 | Pre-training from shape descriptors and on-the-fly synthetic PDE data cuts relative L2 error by up to 47% against training from scratch, with no stored pre-training corpus. 12 |
Agents and tool use
1. DAGent: Evaluate-then-Grow Planning for Deep Research Agents
Area tag: Deep research agents and multi-agent planning
arXiv: 2609.39154, first submitted September 30, 2026. 1
Authors / institutions: Hanwen Liu and Yuanfu Sun at New York University and NYU Shanghai, with Qiaoyu Tan at NYU Shanghai as corresponding author. 13
Status: Accepted at NeurIPS 2026; the submission record reads "Accepted at NeurIPS 2026". 1
Problem: A DAG-based research agent usually builds its whole task graph before running anything, then patches the graph after a failure appears. The paper calls this Plan-then-Patch and argues it commits hardest at the moment the system understands the query least, so later repairs spend compute on branches that should never have been planned. 13
Method: DAGent replaces that schedule with Evaluate-then-Grow. An Orchestrator adds one batch of nodes at a time, conditioning each expansion on structured per-node feedback — outcome, rationale and reliability — returned by nodes that already ran. A hierarchical context layer propagates a compact QueryDoc to downstream nodes by default and keeps the full execution trace available for on-demand recall. Because the graph is append-only, the recorded topology supports DAGRPO, which adds topology-conditioned credit to executor rollouts that feed the final synthesis and a structural-compliance penalty to orchestrator plans that fail validation. 13
Comparison: FlowSearch, with the authors' released implementation run as a controlled baseline, and Flash-Searcher as the strongest training-free baselines, alongside a same-architecture Plan-then-Patch variant and outcome-only GRPO for the reinforcement-learning comparison. 13
Results: At Qwen3-235B-A22B, DAGent leads the strongest open-source baseline by 5.3 / 5.8 / 2.0 points on BrowseComp-Plus / GAIA / xbench-DeepSearch, and the lead repeats across four open-source backbones from four vendors. At Qwen3-32B it reaches 47.3 / 55.3 / 65.0, which is 4.0 / 3.8 / 2.0 above FlowSearch. DAGRPO lifts Qwen3-8B from 40.0 / 46.6 / 60.0 in training-free mode to 49.6 / 53.4 / 65.7, 3.6 / 2.6 / 2.7 above the same-budget GRPO run. Replacing Evaluate-then-Grow with Plan-then-Patch costs 5.3 / 4.8 / 4.0 points with everything else held fixed. 13
Resources: Code is public at the project repository. 13
Loading content card…
Takeaway: A team building multi-step research agents can test the planning change without the reinforcement-learning half, because the ablation isolates Evaluate-then-Grow from DAGRPO. The paper's own limits define the ceiling: DAGent is text-native and plans only over what its tools retrieve, so retrieval quality bounds the result, and the per-node attribution is reported as correlational rather than causal. 13
2. Self-Evolving Coding Rules for AI Coding Agents
Area tag: Coding agents and self-improving harnesses
arXiv: 2610.00650, first submitted September 30, 2026. 2
Authors / institutions: Zhengyuan Jiang, Reachal Wang, Yuepeng Hu, Yupu Wang, Yuqi Jia and Neil Zhenqiang Gong at Duke University. 14
Status: Accepted by NeurIPS 2026. 2
Problem: A coding agent's behaviour is steered by a rules file — persistent instructions about style, libraries and procedure that the agent reads on every task. Teams write that file by hand and then freeze it, so the agent's operating instructions stay at whatever quality someone guessed on the first pass. 14
Method: RuleEvolve keeps a pool of candidate rule sets and improves it in rounds. A mutator module, itself a language model, writes variants of the current candidates; a judge module scores each variant against a verifiable objective — whether the agent's generated code passes the tests — together with a conciseness term; the best candidates replace the pool. The search runs over the rules rather than over model weights, so it applies to a frozen backbone. 14
Comparison: No rules, hand-written rules, published community rules and Prompt-Ops, across two agent frameworks (OpenHands and a custom agent built on the OpenAI SDK), four backbones that include gpt-5-mini and gpt-5.3-codex, and three benchmarks (BigCodeBench, BigCodeBench-Hard and SWE-bench Lite). 14
Results: On BigCodeBench with gpt-5-mini, evolved rules reach a pass rate of 0.46 against 0.31 for manual rules while using nearly 40% fewer tokens; with gpt-5.3-codex the same setup cuts average code length by over 60% without losing pass rate. On SWE-bench Lite with GPT-4.1-mini it reaches 0.340 pass rate, the best of the compared settings, and multi-seed runs hold a pass rate of 0.502 ± 0.026, on par with using no rules at all. 14
Resources: The arXiv record and the paper list no public repository for this work. 2
Takeaway: An engineering team that maintains a coding-rules file can treat this as an automated search over its contents, and the reported win is as much about token cost as about correctness. The paper states its own limits: brevity can cost pass rate on complex tasks, and both the mutator and the judge are language models, so accepted rules inherit their biases. 14
LLM post-training and reasoning
3. Advancing Entropy-Level Credit Assignment in RLVR via Proximal Entropy Policy Optimization
Area tag: Reinforcement-learning post-training and credit assignment
arXiv: 2609.39402, first submitted September 30, 2026. 3
Authors / institutions: Yun Kim and Nojun Kwak at Seoul National University. 15
Status: Accepted at NeurIPS 2026; the record carries the acceptance line with the page count. 3
Problem: Value-free RLVR methods such as GRPO give every token in a rollout the same advantage, even though a few tokens decide whether the derivation holds. Entropy-based fixes weight tokens by how surprising they are, but they rank entropy across the whole training batch, so a token can earn credit for sitting in a hard prompt or near the start of a trajectory rather than for being a real decision point. 15
Method: The paper defines proximal entropy, a token's entropy measured against its own local neighbours rather than against the batch, and shows the quantity is invariant to prompt difficulty and to position. Proximal Entropy Policy Optimization uses that local measure to weight per-token advantages inside an otherwise standard group-relative update. 15
Comparison: GRPO with uniform token weighting, the 80/20 method that updates only the top 20% of tokens by global entropy, and an entropy-augmented advantage, on Qwen3-1.7B, Qwen3-4B and Llama-3.2-3B-Instruct, scored on MATH500, AMC, AIME 2024 and AIME 2025. 15
Results: On Qwen3-4B the method outperforms GRPO by 6.49 points on AIME 2024 and 9.62 points on AMC, and beats 80/20 by 2.27 points in mean accuracy. On Qwen3-1.7B it improves on GRPO by 5.16 points on AIME 2024 and 4.98 points on AMC, with a mean gain of 1.82 points over the entropy-augmented baseline. The diagnostic behind the design is a sharp contrast: global entropy swings by up to 80 points across generation positions and moves from 0.221 on easy prompts to 0.315 on hard ones, while proximal entropy stays flat at about 0.0099 across both. 15
Resources: No project repository is listed on the record; the evaluation uses the public LiveCodeBench implementation. 3
Takeaway: A post-training engineer can drop the local-entropy weighting into an existing group-relative update, and the ablation table is what to read first, because it separates the token-selection change from the training budget. The evaluation stays inside mathematical reasoning with one held-out coding split, and the paper does not compare against value-based methods such as PPO. 15
4. Provable Test-Time Scaling for Beam Search in LLM Reasoning
Area tag: Test-time compute and search theory
arXiv: 2609.38672, first submitted September 29, 2026. 4
Authors / institutions: Qijia He and Yingbin Liang at The Ohio State University; Yu Huang and Yuxin Chen at the University of Pennsylvania; Yuan Cheng at the National University of Singapore. 16
Status: Accepted to NeurIPS 2026. 4
Problem: Beam search that scores partial trajectories with the model's own log-likelihood and only asks an external reward model once a full response is written is a standard test-time method, and its behaviour had no theory: when it prunes the correct answer, and how many samples it needs to keep that answer alive, were both unbounded. 16
Method: The paper models the search through a token-level coverage coefficient C⋆(x) and analyses the standard likelihood-scored, reward-at-the-end recipe. It then proposes Confidence-Filtered Beam Search, which discards low-self-consistency tokens before beam selection, and a self-consistent variant that samples rather than trusts the model's estimated probabilities. Regret is decomposed into reward-model error, base-model capacity and search failure. 16
Comparison: Vanilla beam search, Best-of-N, majority voting and Best-of-Majority; the empirical side runs a synthetic language-model simulator paired with a noisy reward model, plus a pretrained model in the appendix. 16
Results: Vanilla beam search needs at least Ω(C⋆(x)²) samples for the optimal response to survive, where C⋆(x) is the token-level coverage coefficient. Confidence filtering reduces the sufficient coverage dependence from quadratic to nearly linear, Õ(C⋆max), and the self-consistent variant admits a query budget growing only linearly in the horizon L. Sequence-level methods carry coverage terms that grow exponentially with L by comparison. In the simulator, the filtered method's advantage is largest on the hardest instances (per-rollout success probability 0.01) and widens as the horizon grows, where the sequence-level baselines lose accuracy fastest. 16
Resources: The record lists no code repository. 4
Takeaway: The value here is a shape to reason with: the quadratic-to-near-linear change names exactly why a filtered beam outperforms Best-of-N as the horizon grows. The guarantee rests on prefix alignment between likelihood and final reward, and the paper says so, which means the advantage narrows when the reward-optimal prefix is one the likelihood would prune. 16
Language modeling and tokenization
5. Universal Byte-Level Encoding: UTF-8/UTF-16 Routing to Reduce Cross-Script Token-Budget Disparities
Area tag: Tokenization and multilingual efficiency
arXiv: 2610.01984, first submitted October 1, 2026. 5
Authors / institutions: Hyunsik Kim and Youngmoon Jung at Samsung Research. 17
Status: Accepted to NeurIPS 2026. 5
Problem: Byte-level BPE tokenizers cover all of Unicode, but under UTF-8 a script whose characters need three bytes starts from a higher pre-merge cost than English. The paper calls that worst-case cost the encoding floor, and where the vocabulary runs out of learned merges a multibyte character falls back to several byte symbols. Under a fixed vocabulary and context window, CJK, Devanagari and Thai text then spends more tokens per character than English does. 17
Method: Universal Byte-Level Encoding routes each code point to the substrate whose floor is lowest. ASCII stays on its one-byte UTF-8 path; the three-byte BMP scripts are sent through a UTF-16 path that represents them with two Private Use Area symbols, which lowers their floor from three to two. Non-BMP code points keep their four-symbol sequence. The change lives entirely in the tokenizer's byte substrate. 17
Comparison: Standard UTF-8 BBPE, the UTF-16-only BBPE16, and two script-aware tokenizers, MYTE and a script-based BPE, with language-model runs at 182M/32K, 230M/64K, 329M/128K and 1.3B/32K parameters. 17
Results: UBE decodes with a 100% pass rate on FLORES-200 across 204 language–script configurations, and in an exhaustive Unicode 17 audit it recovers every tested raw input, where MYTE recovers 1,138,306 of 1,155,724 inputs and the script-based baseline normalises or filters some scalars away. The cost is 256 extra symbols reserved for the Private Use Area path, an overhead below 1% at vocabulary sizes of 32K or larger. The paper also reports small regressions in a few high-coverage languages and a weaker win at very narrow language coverage. 17
Resources: No repository is listed; the comparison builds on public tokenizers and open models. 5
Takeaway: A team paying a multilingual token tax can evaluate this at the tokenizer stage, before any modelling decision, and the FLORES-200 table is the one to read. The paper is explicit that a new vocabulary makes it unsuitable as a drop-in replacement for an existing checkpoint, so the win applies to a model trained against the new tokenizer. 17
Interpretability and model reliability
6. The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstaining
Area tag: Mechanistic interpretability and hallucination
arXiv: 2609.32964, first submitted September 26, 2026. 6
Authors / institutions: Vy Nguyen, Ziqi Xu, Jeffrey Chan, Estrid He, Feng Xia, Renqiang Luo, Erik Cambria and Xiuzhen Zhang, across RMIT University, Jilin University and Nanyang Technological University. 18
Status: Accepted to the NeurIPS 2026 main conference. 6
Problem: A model asked something it cannot answer often commits to a confident wrong answer instead of abstaining, even when its own internal state carries a signal that the question is unanswerable. Which components turn that signal into the commit-or-abstain decision was unmeasured, so the failure had no component-level account. 18
Method: The authors define a commit-abstain margin Δ(x) from the model's relative likelihood of answering versus abstaining, then use activation patching and component gating over attention heads and MLP sublayers to find the components that move it. The surviving sparse set is the Commit-Abstain Circuit. A lightweight policy trained on the circuit's activations then replaces the model's default threshold for committing, and the same margin doubles as a score for the answerable-versus-unanswerable distinction. 18
Comparison: The model's intrinsic commit-abstain margin, plus prior linear-probe and activation-steering approaches that decode answerability or steer abstention, evaluated on ten models from five families between 3B and 14B and reproduced on 27B–35B models across three benchmarks. 18
Results: The circuit is sparse, covering a median 5.2% of components (from 2.0% in Gemma 12B to 11.8% in Gemma 4B), and it mixes attention heads with MLP sublayers, which make up about 25% of the set. The margin's sign matches the model's actual full-generation behaviour in 97.6% of instances. The policy trained on the circuit's activations improves decision accuracy by 12.2 points over the intrinsic margin, cuts false abstentions by 2.5×, and transfers to unseen benchmarks. On unanswerable questions, 27% cross the commit threshold on average under the intrinsic rule, ranging from 11% for a 9B Qwen model to 41% for a 4B Gemma model. 18
Resources: Code is public at the project repository. 18
Loading content card…
Takeaway: A team that needs its model to abstain on unanswerable input can train a small policy on the circuit's activations instead of re-tuning the whole model, and the AUROC table shows where the margin is already informative. The margin is computed from token-level likelihoods under a fixed answer format, so the method assumes the model can express both an answer and an abstention. 18
7. Mitigating Object Hallucination in Large Vision-Language Models via False Discovery Controlled Visual Data Splitting
Area tag: Multimodal hallucination and error control
arXiv: 2609.38979, first submitted September 30, 2026. 7
Authors / institutions: Chang Liu, Yu Tian and Rui Xie in the School of Data, Mathematical, and Statistical Sciences and the Institute of Artificial Intelligence at the University of Central Florida. 19
Status: Accepted to NeurIPS 2026. 7
Problem: Large vision-language models name objects that are not in the image. Contrastive decoding suppresses some of these errors, but its thresholds are set by hand, and it offers no stated bound on how many of the objects it does report are false. 19
Method: CORAL casts object selection as multiple hypothesis testing. It splits the visual input into overlapping views, scores each candidate object under each view, and applies a mirror-statistic rule from the false-discovery literature to keep the expected share of false discoveries at a chosen level q. The procedure is training-free and reads only token-level scores. 19
Comparison: Contrastive and heuristic hallucination-mitigation baselines on LLaVA-OneVision-7B, Qwen2.5-VL-7B and InternVL3-8B, with further runs on LLaVA-v1.5, InstructBLIP-7B and Qwen-VL-7B, across MSCOCO, A-OKVQA and GQA over 3,000 runs. 19
Results: With the target false discovery rate set to q = 0.1, CORAL holds the measured rate near that level while keeping grounded objects: on one worked example it retains all three real objects, where the baseline reports a fourth, hallucinated object and lands at 25% false discovery. On MME the score rises from 80.8 to 86.7 for LLaVA-OneVision-7B, from 83.5 to 87.9 for Qwen2.5-VL-7B and from 83.4 to 94.5 for InternVL3-8B, a gain of 11.1 points. 19
Resources: The record lists no repository; the authors publish a project page for the work. 7
Takeaway: A team that needs a stated hallucination rate rather than a fixed threshold can set q and read off the trade-off, because the method exposes the false-discovery level as a dial. The method needs token-level logits, so the paper rules out black-box commercial APIs, and it controls errors of object existence rather than relational or attribute errors. 19
Vision and generative models
8. Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE
Area tag: Video diffusion and mixture-of-experts scaling
arXiv: 2609.38140, first submitted September 29, 2026. 8
Authors / institutions: Yu Xu at the University of Chinese Academy of Sciences and ByteDance; Yuxin Zhang, Haotian Yang, Yizhi Wang, Xinwei Huang, Minxuan Lin, Angtian Wang and Chongyang Ma at ByteDance; Xiao Yang at Canva Research; Fan Tang at the University of Science and Technology Beijing. Yuxin Zhang is the corresponding author. 20
Status: Accepted as a spotlight at NeurIPS 2026; the record carries the acceptance line and a project page. 8
Problem: A token-wise mixture-of-experts layer routes each token independently inside one uniform expert pool and regularises expert usage toward balance. Video is spatiotemporally redundant and its semantics are long-tailed, so that design scatters coherent patches across unrelated experts and fragments routing, which shows up as structural distortion in the generated video. 20
Method: SplitMoE splits the expert pool by role. A smaller set of semantic experts learns high-level abstraction and a larger set of generic experts preserves residual visual detail and generative capacity, in a 20/80 split with dual-track routing. Prototype-guided routing, regularised with an attract-and-repel term, lets tokens cluster by semantic attribute instead of by a balancing constraint. The authors upcycle Wan 2.2's low-noise checkpoint, replacing dense feed-forward blocks at even layers from 15 to 35 with sparse MoE layers. 20
Comparison: A dense Wan 2.2 fine-tuned on the same data, a standard Wan 2.2 MoE with 100 generic experts, and same-source ablations that drop prototype guidance or the pull or push terms, all sharing the same checkpoint, data, step count and inference settings. Reported baselines also include CogVideoX-1.5, Mochi, HunyuanVideo, LongCat-Video, LTX-2, Wan 2.2 and OmniWeaving. 20
Results: At A14B activated parameters, SplitMoE reaches 84.47% human fidelity and 69.41% physics on VBench-2, against 73.08% and 54.66% for the dense fine-tune on identical data, and it also leads on creativity (58.46% vs 49.57%) and commonsense (64.89% vs 55.92%). On T2V-CompBench it reaches 84.63% consistency of attributes and 44.01% numeracy, against 77.51% and 36.71% for the dense baseline. Training reaches comparable loss in roughly 70% of the steps. 20
Resources: The record lists a project page at yuci-gpt.github.io/SplitMoE; no source repository is given on the arXiv record. 8
Takeaway: A team scaling a video diffusion model has a controlled ablation here rather than a leaderboard win, because every variant shares one checkpoint and one data pipeline. The paper's own limits are reliance on frozen visual features and the overhead of sparse routing, both of which it leaves to future work. 20
9. Beyond Selection: Token Parameterization for Extreme Visual Token Compression
Area tag: Vision-language efficiency and token compression
arXiv: 2609.35232, first submitted September 28, 2026. 9
Authors / institutions: Rui Zhong, Yu Li, Zheyu Yan and Cheng Zhuo at Zhejiang University. 21
Status: Accepted at NeurIPS 2026 as a spotlight; the record also lists the project repository. 9
Problem: Visual-token compression cuts the cost of a vision-language model, but at extreme budgets pruning tokens breaks the visual grounding that answers depend on, and learned resamplers add parameters, attention cost and training complexity. 21
Method: Braco treats compression as token parameterization and separates two objectives: a basis transform, a two-dimensional DCT over the patch lattice that keeps a low-frequency block, and coordinate organization, which adds input-independent basis-coordinate embeddings plus a budget-dependent orthogonal re-parameterization and learned spatial residual tokens from light pooling. The result is a four-step coder that adds no attention inside the compressor. 21
Comparison: Token-pruning methods and learned resamplers evaluated at matched compression budgets on LLaVA-family and InternVL3.5 vision-encoder-to-LLM pipelines across 23× to 144× compression, on standard single-image multimodal benchmarks. 21
Results: Braco reaches 95.2% accuracy at 23× compression while cutting prefill FLOPs by 84.2–86.7% relative to the uncompressed model, and keeps 90.8% accuracy at 182×, where a comparable prior method reaches 68.9%. At a matched budget it runs about 36% faster end to end than a competing resampler, with 16.6× lower compressor latency and 78.8× fewer compressor FLOPs. 21
Resources: Code is public at the project repository. 9
Loading content card…
Takeaway: A team serving a multimodal model can drop this into an open encoder-to-LLM pipeline and choose a budget from the reported accuracy-cost curve. The paper's own scope is LLaVA-family and InternVL3.5 pipelines on single-image benchmarks, and its coordinate choice is calibrated for the tested encoders, so video, dense localisation or a new backbone would need a small re-calibration sweep. 21
Reinforcement learning and optimization
10. Behavioral Foundation Models for Quality Diversity
Area tag: Reinforcement learning and quality-diversity search
arXiv: 2609.35615, first submitted September 28, 2026. 10
Authors / institutions: Nazim Bendib, Nicolas Perrin-Gilbert and Olivier Sigaud at Sorbonne Université, ISIR, Paris. 22
Status: Accepted at NeurIPS 2026. 10
Problem: Quality-diversity methods search directly in policy parameter space to find a repertoire of diverse, high-performing behaviours. That space is high-dimensional, and with a sparse reward the search has little to select on, so on sparse-navigation and deceptive tasks it collapses. 22
Method: BFM-QD moves the search into the latent space of a behavioral foundation model. The BFM is trained once on an offline dataset and then frozen, and the quality-diversity loop mutates and selects within its 50-dimensional latent space instead of in parameters. The forward-backward structure of the BFM also yields a closed-form, gradient-free improvement operator that approximates a policy-gradient update without training a critic or backpropagating. 22
Comparison: Five parameter-space quality-diversity baselines (MAP-Elites, CMA-ME, PGA-ME, DCRL-ME and DDE-Elites) plus a pretrained variant, across dense locomotion (Walker, HalfCheetah), sparse navigation (AntMaze-medium) and contact-rich manipulation (Cube-single). 22
Results: BFM-QD variants reach the highest normalized QD-score on every task while searching a 50-dimensional latent space, against baselines that search roughly 20,000-parameter policies. On the dense locomotion tasks the margin is modest and the gain is in convergence speed, because the behavioral prior accelerates the search even where the task is easy. On the sparse and deceptive tasks the parameter-space methods collapse toward near-zero coverage while BFM-QD keeps both coverage and fitness. 22
Resources: No code repository is listed on the record. 10
Takeaway: A team that already has a behavioral foundation model can reuse its latent space as the search space for a behavior repertoire, which is the cheaper half of a quality-diversity pipeline. The paper lists three limits: it needs a diverse offline dataset, performance degrades on narrow datasets, and the frozen BFM sets a ceiling on the behaviours the search can reach. 22
11. Adam under Generalized Smoothness with Second-Moment-Type Stochastic Gradients
Area tag: Optimization theory and adaptive methods
arXiv: 2609.37787, first submitted September 29, 2026. 11
Authors / institutions: Ruinan Jin at Mohamed bin Zayed University of Artificial Intelligence; Difei Cheng at Aerospace Information Technology University; Ling Chen and Jun Luo at The Ohio State University; Hao Zhou at JD.com; Youzhi Zhang at the Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences. 23
Status: Accepted at NeurIPS 2026. 11
Problem: Adam stays stable on objectives that violate global smoothness, and the generalized-smoothness analyses that account for this have assumed stochastic gradients are almost surely bounded or sub-Gaussian. Whether Adam converges when the gradients carry only second-moment information was left open in earlier work. 23
Method: The paper analyses a calibrated Adam family under generalized smoothness where the stochastic gradients satisfy a second-moment-type condition and no concentration assumption beyond it. It proves a high-probability convergence bound with explicit dependence on the dimension, the confidence level and the step size, and it separates two regimes by the polynomial growth exponent p of the gradient bound. 23
Comparison: Prior Adam analyses under generalized smoothness that require bounded or sub-Gaussian gradients, and the earlier open question those analyses left. 23
Results: For the regime 1 ≤ p < 2 the paper constructs a fixed class of one-dimensional hard instances, containing a deterministic quadratic and a countable staircase family, on which the calibrated Adam family has a worst-case expected squared-stationarity lower bound of Ω(T^−1/3), uniform over deterministic schedules; for 0 ≤ p < 1 it proves a matching high-probability upper bound. The bound's leading constants do not depend on the dimension, and the step-size height is independent of the confidence parameter. 23
Resources: No code repository is listed on the record. 11
Takeaway: The result removes a tail assumption from the theory of Adam and gives a lower bound that tells you the rate cannot be improved in general, so it is a reference point rather than a new algorithm. The construction that yields the lower bound is one-dimensional and adversarial, which is where the guarantee is tightest. 23
Scientific machine learning
12. Geometry Meets Physics: Data-Efficient Pre-Training for Unstructured Neural PDE Solvers
Area tag: Scientific machine learning and neural PDE surrogates
arXiv: 2610.03363, first submitted October 2, 2026. 12
Authors / institutions: Luis Medrano-Navarro, Giacomo Baldan, Qiang Liu, Benjamin Holzschuh and Nils Thuerey at the Technical University of Munich; Jan Hagnberger and Mathias Niepert at the University of Stuttgart. 24
Status: Accepted to NeurIPS 2026. 12
Problem: Neural surrogates for partial differential equations on unstructured three-dimensional geometry generalise poorly and need large labelled simulation datasets. Pre-training on a big corpus of related dynamics helps, but generating and storing that corpus is expensive in compute and storage, so the strategy trades one cost for another. 24
Method: The framework removes the pre-computed corpus. For steady-state problems it learns a representation from intrinsic shape descriptors of the three-dimensional domain, as a form of geometric pre-training. For transient problems it generates synthetic PDE data online during pre-training, so nothing is stored, and it couples the geometry and physics of the problem instead of treating them separately. A target model is then fine-tuned on the task at hand. 24
Comparison: Training from scratch for both steady-state and transient regimes, and a geometric pre-training baseline, with backbones that include AB-UPT and other neural-operator models on the DrivAerML automotive benchmark and a wing dataset. 24
Results: On the steady-state tasks, pre-training raises accuracy most in the sparsest data regimes, with a maximum 47% improvement at 16 training samples, and fine-tuning on 16 samples beats training from scratch on 32. On the transient tasks the paper reports consistent gains of around 30% in relative terms across the three benchmarks, with larger absolute gains in the smaller-data regimes. The pre-training data are generated on the fly, so none of it is stored. 24
Resources: The record lists no code repository; the paper evaluates on the public DrivAerML benchmark and the authors' own released SHIFT-Crash and Wing datasets. 12
Takeaway: A team training a PDE surrogate on a low-data geometry can use the online data generation to pre-train without first paying for a corpus, and the gain is largest where labelled simulations are scarce. The gain is largest where labelled simulations are scarcest, which is the regime the paper targets, and the transient runs show a steadier 30% relative improvement. 24
What to open first
- Building or automating agents: DAGent (2609.39154) if your agent plans a multi-step research task, because the ablation isolates the planning change from the reinforcement learning; RuleEvolve (2610.00650) if you maintain a coding-rules file and want its contents searched rather than written by hand. 1
- Post-training or serving reasoning models: PEPO (2609.39402) for the token-level credit-assignment knob; the beam-search paper (2609.38672) if you run test-time search and want to know why filtering beats Best-of-N as the horizon grows. 3
- Running a multilingual model: UBE (2610.01984) if your token cost differs by script and you are still choosing a tokenizer. 5
- Diagnosing a failure mode: the Commit-Abstain Circuit (2609.32964) for why a model commits to an answer it should refuse, and CORAL (2609.38979) if you need a stated false-discovery rate on object hallucination rather than a fixed threshold. 6
- Scaling video or cutting multimodal cost: SplitMoE (2609.38140) for video-diffusion scaling under a controlled ablation; Braco (2609.35232) if prefill compute is your bill and you can afford a small accuracy drop. 8
- Working on reinforcement learning or optimizers: BFM-QD (2609.35615) for exploration in sparse-reward tasks, and the Adam paper (2609.37787) if you need the convergence assumption spelled out. 10
- Training scientific surrogates: Geometry Meets Physics (2610.03363) if your labelled simulation data are scarce and your geometry is unstructured. 12
References
- 1DAGent arXiv record
arxiv.org
- 2RuleEvolve arXiv record
arxiv.org
- 3PEPO arXiv record
arxiv.org
- 4Beam search arXiv record
arxiv.org
- 5UBE arXiv record
arxiv.org
- 6Commit-Abstain arXiv record
arxiv.org
- 7CORAL arXiv record
arxiv.org
- 8SplitMoE arXiv record
arxiv.org
- 9Braco arXiv record
arxiv.org
- 10BFM-QD arXiv record
arxiv.org
- 11Adam arXiv record
arxiv.org
- 12GMP arXiv record
arxiv.org
- 13DAGent paper
arxiv.org
- 14RuleEvolve paper
arxiv.org
- 15PEPO paper
arxiv.org
- 16Beam search paper
arxiv.org
- 17UBE paper
arxiv.org
- 18Commit-Abstain paper
arxiv.org
- 19CORAL paper
arxiv.org
- 20SplitMoE paper
arxiv.org
- 21Braco paper
arxiv.org
- 22BFM-QD paper
arxiv.org
- 23Adam paper
arxiv.org
- 24GMP paper
arxiv.org
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.