Tuesday, October 6, 2026

Coding agents can write numerical solvers that run, yet fail on hard cases because of conditions like ill-conditioning or stiffness, and runtime feedback rarely shows the cause. ADSD (Auto-Diagnosis and Skill Discovery) first works out why a solver performs poorly and only then searches the literature for a matching method, which it packages as a reusable solver skill without updating any model weights.

• How it works: a frozen baseline solver is run on public test cases, and claims drawn from a curated paper database are turned into diagnostic probes; hypotheses are kept only if they hold on over 90% of a separate verification pool, and each surviving diagnosis then guides retrieval of methods that are theory-audited, implemented, checked empirically and given to a fresh agent to rebuild the solver.
• Power systems: on GOC-500 power flow, Qwen3.8-2.4T's mean error falls about 71×, from 1.0e-9 to 1.4e-11, with no loss of speed, and its pass rate on near-loadability AC-OPF edge cases rises from 14.4% to 98.4%; skills learned on the N2 topology also transfer to the unseen N1 and N grids.
• ODEs and PDEs: on stiff systems-biology ODEs, GPT-5.5's pass rate climbs from 70.1% to 90.6% while median trajectory error drops about 24×, and on high-contrast ConDiff diffusion problems GPT-5.5's operator calls fall from 215 to 71 and its runtime from 7.3 to 2.8.
• Ablations: asking the agent to improve its own solver without diagnosis gives inconsistent results (error fell in four of nine comparisons and rose in five), and five recursive rounds of self-revision never closed the gap to ADSD.
• Limitations: several gains are modest (for example, Claude-Sonnet-5 on IEEE-118), improvement on close-to-infeasible cases is limited by the problem itself, the pipeline depends on a pre-built domain literature database, and off-the-shelf solvers are excluded, which may understate what practical baselines can achieve.

arXiv:2610.03872 PDF

Picking a harness for each query (which agents, instructions, tools and wiring a multi-agent system uses) has usually meant trying candidate designs at inference time or building them by hand. SHIFT avoids this with a small local "architect" model that predicts how useful a harness will be, trading accuracy against cost, without running it. Monte Carlo tree search uses these predictions to build a harness per query, and only the chosen harness is executed.

• The architect is a frozen Gemma 4 E2B with LoRA adapters plus policy and value heads; it searches over 40 actions (adding agents, adding instructions, granting tools), and its value head is trained on a reward that credits task success and penalizes tokens, latency, tools and agents.
• Across 9,193 tasks in six benchmarks with a Gemini 3.5 Flash executor, SHIFT-search reaches 79.9% mean accuracy, beating the strongest of 17 baselines (Trace, 72.7%) by 7.2 points, with the biggest gains on OfficeQA (59.3% vs 29.3%) and GAIA (68.6% vs 53.6%).
• The cheaper SHIFT-value variant picks the candidate with the highest predicted utility and reaches 74.5%, still above every baseline, while using 32% fewer execution tokens than Trace; building a harness takes under a second on one H100.
• On HotpotQA, choosing agents, instructions and tools together (69.4%) beats tools only (67.9%) and instructions only (60.3%); an architect trained only on GSM8K transfers to MATH-500 (83.6% vs 81.2% for a fixed single agent).
• SHIFT-search uses roughly twice as many execution tokens as Trace, gains little on short-horizon tasks where the executor model is the bottleneck, and needs a separately trained architect for each benchmark, which adds upfront compute.

arXiv:2610.04137 PDF

Diffusion language models decode many tokens in parallel but still trail autoregressive models of the same size on quality, and the authors attribute this to a computation–difficulty mismatch: every masked position gets the same depth at each denoising step, regardless of how hard it is to predict. ALoDLM fixes this with token-adaptive latent recurrence: within each step, easy tokens commit early as discrete context, while hard tokens keep refining their hidden states through more passes of a looped middle block.

• A Qwen3 backbone is split into a prelude, a recurrent core and a coda; on each inner pass, tokens whose entropy falls below a threshold τ are sampled and fed back as embeddings, while a learned exit gate with a cumulative halting threshold q decides when to stop refining the remaining tokens, up to a maximum depth of K = 4.
• Training treats each token's exit depth as a latent variable and optimizes a conditional NELBO with an unbiased score-function gradient estimator, adding supervision at intermediate depths and a first-pass control variate to reduce variance; both models are converted by direct SFT on just 5B tokens, with no continued pretraining.
• Across eleven reasoning, math and coding benchmarks, ALoDLM averages 65.5 at 1.7B and 80.3 at 8B, beating the Qwen3 autoregressive baselines (63.8 and 78.5) and the strongest diffusion baseline, WeDLM-8B, by 5.2 points, with the lead on ten of eleven tasks; these quality numbers were measured with every model decoding one token per step.
• On GSM8K, ALoDLM-8B reaches about 2.7× the throughput of vLLM-served Qwen3-8B at comparable or higher accuracy; at a matched 93.25% accuracy it beats WeDLM-8B with 612.4 versus 564.2 tokens/s and 133.5 versus 154.6 GFLOPs per token, but WeDLM stays ahead at the fastest operating points, above roughly 650 tokens/s.
• Raising q at inference increases the average loops per token from 1.6 to 2.34 and the 11-benchmark average from 77.9% to 79.1%, so recurrence depth works as a test-time scaling knob; the main limitations are a longer time to first token, because prefill builds the KV cache at full recurrent depth, and input-dependent speed that could shrink or reverse the speed advantage on domains underrepresented in training.

arXiv:2610.04198 PDF

Even accurate, task-relevant memories can push LLM agents to side with a user's past views: in a controlled test, adding such memories cut accuracy from 96.0% to 77.3%. MemAdapter therefore leaves memory filtering alone and instead decides after retrieval how much each memory should be allowed to influence the current task.

• The method runs three prompting stages: Counterfactual Induction tests each memory against hypothetical tasks to find where its influence stops being appropriate, Context-Aware Reflection turns those findings into a per-memory usage instruction for the actual request, and Evidence-Based Reasoning generates the answer along with an internal trace that checks every use of memory stays within its instruction.
• With DeepSeek-V4-Flash across five memory systems, it gave the best scores on both memory-use dimensions of MemSyco-Bench every time, lifting "When to Use Memory" accuracy from 32.11% to 88.78% on LightMem and from 44.35% to 87.89% on Mem0, and it cut the PersistBench sycophancy failure rate by 21–31 points on all five systems.
• The gains carry over to other backbones: on Qwen3-8B, average MemSyco-Bench accuracy rose from roughly 25–38% to 80–87%, while gains on GPT-5.6-sol were smaller (+3 to +40 points) and largest on the weaker memory systems.
• Ablations show Counterfactual Induction alone does most of the work (average accuracy 70.25 to 84.26 on NaiveRAG), and adding evidence-grounded generation brings the full system to 89.94.
• There are trade-offs: the failure rate on beneficial memories rises for some systems (3% to 9% on MemoryBank), personalized-memory accuracy drops by up to 9 points under GPT-5.6-sol, and over the strongest baseline on A-MEM it leads by only 0.44 points on "When to Use Memory".

arXiv:2610.05162 PDF

Deployed LLM agents see a stream of related long-horizon tasks, yet in-context memory methods leave the policy frozen. Naively fine-tuning on single self-generated trajectories destabilizes the policy. ASCENT instead does online test-time training: it distills each verifier-accepted trajectory into persistent LoRA fast weights. The teacher is the frozen initial model, given the successful trajectory as hindsight privileged information.

• The agent gets one attempt per task in a single pass over the stream, and updates happen only on verified successes; the frozen teacher, shown the full trajectory with invalid-action turns filtered out, gives full next-token distributions that the student matches via forward KL at its own generated prefixes.
• Direct baselines such as Online RFT, RFT+KL, REINFORCE and REINFORCE++ drift from the base model, produce more and more invalid actions, and end up below the untrained model's success rate by the second half of the stream, which motivates soft-target self-distillation over hard-target imitation.
• On ALFWorld seen tasks, ASCENT lifts Qwen3.5-4B from 46.4% to 69.5% and Qwen3.5-9B from 55.0% to 77.4%, beating the best online memory baseline (MemP, 63.6% and 67.9%) and cutting mean turns from 32.6 to 21.0 at 9B.
• On WebShop, strict success rises from about 17% to roughly 42% at both scales, and mean turns drop from 39.5 to 17.8 at 4B; gains also hold when the adapted weights are transferred frozen to held-out ALFWorld scenes, and further online training on those scenes adds more at 4B.
• Limitations: the method needs full teacher logits, so it works only with open-weight models; it gates on sparse episode-level outcomes with no per-turn credit; and the teacher's privileged context includes the student's own response text, which may make its hindsight less independent.

arXiv:2610.05303 PDF

Kandinsky 6.0 Video generates 5-second video clips with synchronized 44 kHz audio, including lip-sync, from either a text prompt or a starting image. It does this by attaching a newly trained audio stream to the pretrained Kandinsky 5.0 video model through bidirectional cross-attention, giving a dual-stream CrossDiT. It comes in two sizes, Lite (3B) and Pro (29B), and the code, weights and diffusers integration are released under the MIT license.

• The audio stream is first trained from scratch on 40M audio tracks while the video stream continues from Kandinsky 5.0; the two are then trained jointly on 7M paired audio-video clips filtered for lip-sync quality, with the noise level sampled independently for each modality.
• After pretraining, the authors fine-tune a separate model on each of 11 content domains and average their weights into a "model soup," apply reinforcement learning adapted from OmniNFT with eight video, audio and sync rewards, and distill the result with π-Flow plus an adversarial refinement stage (Sim-LADD) down to 10 sampling steps.
• Reinforcement learning cuts Pro's speech word error rate from 0.235 to 0.124 (−47%) and Lite's from 0.179 to 0.127; because the Whisper model used for that measurement is also one of the training rewards, the authors confirm the gain on Harvard sentences, where it drops from 0.251 to 0.180.
• In side-by-side human evaluation Pro clearly beats Kandinsky 5.0 Video Pro and is competitive with proprietary systems, especially in speech quality, and it outperforms the open-source LTX 2.5 on most VABench metrics.
• The main limits are that clips are capped at 5 seconds and generated at SD resolution, so Full-HD output depends on a separate 1.4B-parameter super-resolution model, and reinforcement learning was run only at 512×768.

arXiv:2610.05608 PDF

Masked diffusion language models re-predict each answer token at every denoising step before committing it, so an adversary who can hook the residual stream can watch how likely the target answer is and adjust the intervention mid-generation. The attack exploits this with a proportional-integral (PI) feedback controller that sets the strength of a fixed bias-steering vector at each step. It pushes a frozen dLLM toward a demographic answer the adversary picks, without changing the weights, the prompt, or the sampler.

• How it works: a single unit direction is fitted offline at block 14 as the mean activation difference between completions naming the target group and completions naming the comparator group. At each denoising step, a PI controller (setpoint 0.9, K_p=3, K_i=0.1, α capped at 6) reads the probability of the target answer letter and sets the strength at which the direction is added to all 32 blocks of LLaDA-8B-Instruct on the next pass.
• Main results: on ambiguous BBQ questions, where abstention is the correct answer, the target–comparator gap rises from 1.8 to 16.7 pp, while no fixed-strength baseline (CAA, ActAdd, ITI-C, AurA, and others) exceeds 4.6. On SocialStigmaQA, stigmatizing answers rise from 17.6% to 58.1%, and the "Old" age target shifts by up to 37.2 pp, with each attack taking about 40 minutes on one GPU.
• Feedback drives the gain: a constant strength matched to the controller's average (α=3.28) reaches a gap of only 3.5 and corrupts 20.8% of outputs, against 7.8% under feedback. Even a per-item constant chosen with hindsight reaches just 8.4, consistent with the analysis that no single strength suits every example.
• Limitations: the attack costs coherence (7.8% invalid outputs versus 0% unsteered), and on the Asian, Latino, and White targets invalid outputs exceed 15%, so the authors count those runs as failures. Transfer to other models is mixed: the gain on Dream-7B is small (+4.2 pp), and a tuned CAA constant beats feedback on LLaDA-MoE by 2.3 pp.
• Scope: the attack was only tested on multiple-choice answers read from a single letter probability, and the sign-inverted version used as a defense was only tested where the unsteered bias was already near zero. The authors argue that bias audits must examine the serving stack, not just the checkpoint.

arXiv:2610.05894 PDF

World Action Models built on a Mixture-of-Transformers (MoT) design still need over 300 ms per inference call on an H100. Two things cause this: the action expert denoises over many steps, and it cannot start until the video expert has finished its whole forward pass. RealtimeWAM fixes both through post-training, distilling the action expert down to a single denoising step and pipelining the two experts block by block so they run at the same time.

• Teacher-Anchored Consistency Distillation (TACD) adds a second loss to standard consistency distillation that pulls the student's velocity toward the average velocity leading to the endpoint of a frozen teacher's 10-step rollout, since a small local consistency error alone doesn't guarantee an accurate final action; only rank-128 LoRA weights in the action expert are trained, and the video expert stays frozen.
• Cross-Expert Wavefront Pipelining (CEWP) runs the video and action experts on separate CUDA streams, where the video expert publishes each block's KV cache right after projection and the action expert waits only immediately before the attention step that uses it.
• On RoboTwin 2.0, the one-step model reaches 92.64% on the Faster-WAM base versus 92.93% for the 10-step original, and 90.84% on the Fast-WAM base versus 91.51%; this beats one-step Flash-WAM (81.41%) and other distillation baselines such as DMD (88.22%), while LIBERO scores of 97.0% and 99.0% match or slightly exceed the originals.
• With all optimizations on, latency falls from 299.7 ms to 12.2 ms (24.55×) for Fast-WAM and from 218.9 ms to 16.1 ms (13.56×) for Faster-WAM on an H100, though part of that comes from CUDA Graphs and LightX2V kernels, as CEWP by itself adds only 1.34× and 1.17×.
• The method does not apply to shared-backbone WAMs such as DreamZero, the teacher rollouts add about 50% to training time, and the teacher-anchored loss improves only modestly over plain consistency distillation in the ablation (89.85% to 90.84% on RoboTwin).

arXiv:2610.06617 PDF

Looped language models pay for every recurrence: activations stored in training, key-value (KV) cache in decoding, compute in prefill and replays during reinforcement learning (RL). Once recurrent states sit near fixed points, the endpoint can stand in for the whole trajectory, so the authors improve the two training components that shape those fixed points, the depth prior and input injection.

• The endpoint enables four shortcuts: truncated backpropagation through the last few loops, decoding that caches only each token's terminal KV, a distilled student that prefills in place of the loops, and RL updates computed from saved rollout states. Caching only terminal KV cuts the cache 3× at five recurrences, but under fixed-depth training with full backpropagation it drops GSM8K accuracy from 50.6 to 21.2 at 1.6B parameters.
• The learned depth prior replaces Huginn's fixed Poisson–log-normal prior with a categorical distribution over depths 1–64, trained by policy gradient on an inverse-perplexity reward, with an entropy bonus and a mean-depth penalty; it lowers validation perplexity by 0.7–1.8% for at most 1.6% more FLOPs, and at 1.6B it reaches a 60.0 downstream average with a 3× smaller KV cache (the entropy term matters, since GSM8K falls from 47.92 to 36.47 without it).
• OrthoInj projects the carried-over state off the injected input, so each recurrence receives the input at the same strength, and it lowers perplexity by 0.3–1.4% relative to Parcae's injection from 100M to 1.6B, though the downstream gain at 1.6B is marginal (59.26 vs 59.22) and GSM8K is lower there.
• Reusing saved rollout states with a truncated Neumann-series gradient makes the scoring-and-backward step of Dr. GRPO updates ~2× faster (1.99× on GSM8K, 2.02× on MBPP+) while staying within 1.6 points of full-backpropagation pass@1, and distilled prefill runs 1.5–1.8× faster on 8K prompts but trails the teacher's downstream average by 1.0–4.5 points, a gap that grows with scale.
• Experiments stop at 1.6B parameters with one seed per configuration, and the ablations do not cover the full design space.

arXiv:2610.06833 PDF

Fixing just the first one or two tokens of a base model's response can unlock reasoning on par with RL-trained counterparts, and the authors trace this to associations learned from training data. By editing mid-training data and retraining, they show these "token cues" behave like conditioned stimuli: an arbitrary word can be turned into a reasoning trigger, and an existing trigger can be redirected.

• Cues are found without reference answers: beam search proposes high-probability two-token openings, and the one whose sampled answers agree most (lowest answer entropy) is selected; on Olmo-3-7B this pick lands within 1.1 points of the best of 500 tested openings.
• Prefilling .\n\nOkay raises Olmo-3-7B's MATH-500 pass@1 from 42% to 78% (RL-Zero reaches 75%), and prefilling Alright, lifts Qwen3-14B from 72% to 87%, matching its GRPO-trained model, with gains carrying over to GSM8K, AMC 23, and AIME 2024; the improvement holds at matched token budgets, so it is not just longer generation.
• RL's changes to the next-token distribution are concentrated in the first two output positions, raising the probability of the cue from 0.14 to 0.65 for Olmo and from 0.04 to 0.58 for Qwen, and prefilling the cue before any updates matches roughly 100 (Olmo) and 50 (Qwen) GRPO steps.
• Replacing "okay" with "chicken" in a 10B-token mid-training mix raises .\n\nChicken-cued MATH-500 accuracy from 2.4% to 37.2%, while pairing "Okay," with questions drops the original cue to 0.2%, and swapping "step by step" for "duck duck goose" makes "Think duck duck goose" about as effective as "Think step by step" (17% vs. 15%).
• Cues also change safety behavior on XSTest: Okay, produces harmful responses to unsafe prompts 42.4% of the time versus 0.2% for .\n\nOkay; the authors note the findings cover only R1-Zero-style RL applied directly to a base model, and they found no effective cue for Llama-3.1-8B.

arXiv:2610.06851 PDF