Weekly Recap

Weight updates leave readable textual imprints: decoding parameter shifts into natural language enables direct, gradient-based behavioral steering without any downstream fine-tuning data.

Internal activation steering cannot displace standard alignment at scale, but lightweight internal probes paired with runtime interventions rescue models from the safety degradation caused by downstream fine-tuning at near-zero marginal inference cost.

Latent communication between LLMs was largely an illusion of interface shortcutting until Draft-KV, which routes genuine drafting KV caches to let a frozen 0.5B model leap from 37% to 78% MMLU-Redux accuracy by tapping an 8B model's latent thoughts.

Turning text into images cuts reranking token overhead by up to 50% and yields a 1.70× throughput boost—all while outperforming standard sub-4B text rerankers across BEIR.

Uniform layer looping wastes latent compute on trivial tokens, but dynamically routing iteration depth beats standard transformers on AIME with a 53% steeper test-time scaling slope.

Weight updates leave readable textual imprints: decoding parameter shifts into natural language enables direct, gradient-based behavioral steering without any downstream fine-tuning data.

Internal activation steering cannot displace standard alignment at scale, but lightweight internal probes paired with runtime interventions rescue models from the safety degradation caused by downstream fine-tuning at near-zero marginal inference cost.

Latent communication between LLMs was largely an illusion of interface shortcutting until Draft-KV, which routes genuine drafting KV caches to let a frozen 0.5B model leap from 37% to 78% MMLU-Redux accuracy by tapping an 8B model's latent thoughts.

Turning text into images cuts reranking token overhead by up to 50% and yields a 1.70× throughput boost—all while outperforming standard sub-4B text rerankers across BEIR.

Uniform layer looping wastes latent compute on trivial tokens, but dynamically routing iteration depth beats standard transformers on AIME with a 53% steeper test-time scaling slope.

Static parallelism layouts leave massive post-training efficiency on the table, but dynamically re-slicing distributed model states across GPUs cuts RL step latency by 28% with less than 0.1% transition overhead.

Halting LLM debate consensus-drift to stop groupthink can backfire severely: a probe-gated freeze stopped 29 collapses into wrong answers at the direct cost of sacrificing 108 valid self-corrections.

Treating on-policy distillation as value-based RL slashes rollout compute by 75% while actively preserving the generation diversity needed for high-k reasoning.

Distilling multi-billion-parameter multi-vector retrievers no longer requires caching terabytes of document tokens: entropic optimal transport allows a lightweight 149M text model to capture 95% of VLM teacher performance at 26x query speed using queries alone.

Retaining even a single structural relative slashes post-deletion retraining loss by $8\times$, revealing that target accuracy after unlearning is often driven by intrinsic dataset redundancy rather than unlearning failure.

Multi-agent LoRA systems can slash time-to-first-token by 3.1× without sacrificing role specialization by decoupling shared base KV caches from precomputed low-rank adapter updates.

LLM agent token consumption swings by over an order of magnitude on identical tasks, but dynamically tracking context inflation across execution segments slashes budget waste by 21.3% without making a single extra LLM call.

Downstream reinforcement learning easily restores stolen reasoning performance from degraded API outputs, exposing critical vulnerabilities in current frontier model distillation defenses.

Decoupling structured semantics from frame-level actions and distilling autoregressive rollouts via shared, compressed memory tokens solves the long-standing trade-off between real-time responsiveness and multi-step drift in interactive world models.

Transformer row representations fail to form an autonomous thermodynamic critical class during training, meaning popular physical scaling analogies collapse unless explicit optimizer memory and readout error budgets are accounted for.

Extra test-time "thinking" cannot recover missing parametric facts, but teaching small reasoning models to recognize their own knowledge boundaries and selectively query larger models lets a 4B model outperform a 14B baseline at nearly a third of the serving cost.

Personalized alignment is fundamentally a candidate-matching problem rather than a generator bottleneck: million-parameter MLPs tapping internal LLM embeddings beat monolithic, billion-parameter reward models across the board at 10,000× lower latency.

Frontier pretraining runs routinely stake tens of millions of dollars on scaling laws, yet the empirical pipelines used to fit and extrapolate those curves have never been systematically benchmarked until now.

Downstream classifier accuracy can actually improve when neural compressors intentionally distort reconstruction distributions away from mismatched source domains toward a target distribution.

Human hand demonstrations don't need to be tracked or imitated to solve dexterous manipulation—they just need to seed the simulation's reset distribution.

Lazy-loading agent skills on demand doesn't just slash context window costs—it actively sharpens tool retrieval quality, all while incurring only a negligible hit to latency.

Real-time multi-stream interaction does not require bespoke model architectures: off-the-shelf LLMs can natively interleave perception and action in dynamic environments when configured with overlapping-memory inference coroutines.

Collecting more errors doesn't guarantee better models: targeting only the high-impact failure modes signaled by a network's own internal representations outperforms standard rarity-based active learning while saving annotation budget.

Standard mechanistic circuits can mimic over 97% of a model's correct outputs while missing up to 89% of its errors, exposing that common ablation validations reward functional mimicry rather than true causal mechanisms.

Contextual bandits don't actually need reward estimation—exploiting the elicitation geometry of shared monotone links lets algorithms target the decision boundary directly to secure logarithmic regret under severe functional misspecification.

NVIDIA's hand-crafted GEMM heuristics leave massive performance on the table: grounding learned kernel selectors in static hardware-behavior estimates cuts selection regret by up to 64.2% without running an autotuning benchmark.

Neural PDE solvers no longer need retraining for new boundary conditions: learning domain geometry through its harmonic measure enables zero-shot Poisson solutions via simple re-integration against a single transformer kernel.

Modifying executable agent harnesses rather than model weights allows test-time adaptation to generalize across unseen reasoning tasks without a single gradient step.

Transformers actively protect their massive activation outliers through an upstream "read-blindness," where layers continually write to extreme residual coordinates while systematically refusing to read them.

Pretrained scientific foundation models can overcome massive chemical and instrumentation shifts entirely at test-time by using reference-guided retrieval to recalibrate latent predictions on the fly without query supervision.

Shared workspace files can turn stateful LLM assistants into self-replicating malware vectors, silently compromising up to 80% of independent agents across eight consecutive hops via persistent memory.

Moving quality-diversity optimization from raw policy parameters into the latent space of a behavioral foundation model rescues evolutionary search from total collapse in sparse, deceptive environments while eliminating critic training entirely.

Multimodal models no longer need to rely on entangled black-box fusion: invertible flows and low-rank latent priors can explicitly isolate shared task signals from emergent cross-modal synergy.

Robots can master delicate tool manipulation directly from unpaired human videos, achieving a 73% performance leap over SOTA by anchoring control in object-centric 3D pose priors rather than raw end-to-end pixels.

Aggregate benchmark scores actively mislead in quantum computing, failing to predict individual model competence across more than a third of tasks like circuit compilation and quantum error correction.

Categorical sampling cripples discrete diffusion by discarding intermediate uncertainty—diffusing directly over the probability simplex fixes this information collapse, enabling an 8-step distilled model to beat 128-step discrete baselines on GSM8K.

Geometric deep learning has long been fragmented by bespoke, manifold-specific architectures, but a single unified formulation can now systematically derive valid linear and convolutional layers across ten distinct Riemannian geometries.

Shrinking an LLM's residual stream based on activation variance alone discards the wrong directions—weighting pruned subspaces by downstream output sensitivity dramatically improves compressed model performance without sacrificing closed-form computational efficiency.

Non-CUDA hardware no longer requires costly post-training pipelines: conditioning memory rewards strictly on adopted code enables zero-shot LLM agents to synthesize NPU kernels that outperform vendor libraries by up to 44x.

Curbing a model's sycophancy through internal steering does not strengthen baseline safety refusals, though it recovers up to 95% of refusal robustness when users actively push back.

An adversary can permanently cripple deep ReLU networks without ever altering model parameters or labels, weaponizing the "dying neuron" pathology through subtle batch reordering and gradient-inversion poisoning.

Just two incoming connections per agent completely break the finite-depth impossibility of single-parent aggregation, recovering the globally optimal predictor across arbitrary distributions in $O(d \log d)$ depth.

Billion-parameter LLMs no longer require global backpropagation to pretrain competitively: sharing a single delayed readout across decoupled modules matches standard zero-shot accuracy while yielding up to 1.44x pipeline throughput.

In regimes where high-accuracy gradients are brutally expensive to compute, minimizing a convex objective costs asymptotically no more than evaluating its gradient just once.

Runtime scaffolding beats model scale in scientific discovery: structured hypothesis tracking and trajectory branching push DeepSeek-v4-flash from 20.7% to 73.0% accuracy on blind symbolic regression, matching GPT-5.5 without relying on semantic domain clues.

Sub-3-bit LLM inference finally sheds its dequantization bottleneck: decoding Leech lattices on-the-fly via a 16 KiB Golay trellis yields over 113 tokens per second while remaining within 2.5 MMLU points of 4-bit AWQ.

Domain-specific agent evaluation no longer requires manual ground-truth curation: autonomous research and verification loops constrained by reusable expert rules yield grading rubrics that professional analysts prefer over expert baselines.

Static problem pools cause LLM reasoning post-training to stall as models saturate the dataset, but dynamically synthesizing frontier tasks via regret-guided procedural generation keeps learning signals saturated for sustained gains.

Adam's complex adaptive behavior boils down to an invariant, heavy-tailed moment ratio that enables zero-overhead 4-bit optimizer compression and formally unifies Adam with sign-based momentum.

Linear attention models do not need larger hidden states to fix their memory bottleneck; simply assigning attention heads to distinct temporal resolutions unlocks up to an 18.9% boost on associative recall.

Even when provided with only target column metadata, LLMs outperform specialized AutoML heuristics like AutoGluon at inferring machine learning task types (0.98 vs. 0.93 macro F1), unlocking fully autonomous pipeline configuration.

Hidden dynamical systems unidentifiable under partial excitation can be fully recovered using transient, unknown local interventions—bypassing classical state-reachability limits without ever measuring the perturbation coefficients.

Parallel LLM scientific agents collaborate far more effectively when sharing raw evidence rather than hypotheses, unlocking a 9.5-point discovery quality gain and nearly doubling exploration speed without succumbing to premature consensus.

Accurate latent reconstruction is a red herring for diffusion-based reasoning; asynchronous denoising over multi-layer teacher representations allows a compact 638M model to solve complex math and code without autoregressive rollouts.

Standard post-training quietly collapses LLM discourse into an ideological monoculture, but domain-specialized models can recover the diverse perspectives that broad alignment erases.

A 13-parameter statistical count model with Kneser-Ney smoothing outperforms deep temporal graph networks across multiple standard benchmarks without learning a single node embedding.

Interleaving expensive Muon spectral steps with cheap Lion sign descent not only cuts distributed training wall-clock time by 33%, but actually beats full Muon and AdamW on final perplexity.

Aligning ternary weight quantization directly with SIMD ciphertext packing layouts eliminates up to 80% of expensive FHE multiplications, unlocking sub-1% accuracy loss speedups across both ViTs and BERT.

Human hair priors can substitute for non-existent animal fur datasets, enabling dense, editable 3D animal grooming with an order-of-magnitude faster optimization.

Instead of spending compute fitting bespoke generative models for every dataset, a single feedforward pass can now compile empirical samples directly into inspectable, executable generative programs.

Multi-round active simulation turns standard neural density estimators into high-precision, test-time-specialized solvers for notoriously ill-posed physical inverse problems.

Frontier coding agents can now write 100% functionally correct GPU physics simulations, yet they match human-expert performance on just 22% of tasks—revealing a stark capability gap between generating working code and producing performant systems.

Long-horizon agent RL frequently stalls because fixed rewards cannot track shifting bottlenecks—co-evolving evaluation rubrics with targeted exploration skills directly breaks the repetitive failure loops of current policies.

Chasing large token-level teacher-student gaps in agent distillation is largely misguided: local divergence fails to predict task completion, whereas unpenalized minor deviations often trigger catastrophic multi-turn failure.

Standard early-exit architectures collapse completely outside their pre-selected exits, but stochastic prefix supervision transforms a normal Transformer into a valid, deployable language model at literally every intermediate layer depth.

Tool-using agents falsely claim success in nearly a quarter of tool failure cases, but enforcing structured evidence contracts slashes these deceptive reporting errors to under 1% while maximizing task recovery.

Standard RLVR observations are mathematically insufficient to isolate or fix verifier exploitation, proving that reasoning models cannot unlearn rewarded errors without external audit signals.

Monolithic n-gram memory tables waste capacity and fail on polysemous context, but decomposing lookups into a shared basis dictionary with per-component gating unlocks scalable, context-aware parameter expansion.

Image models do not need noisy vision-language reward models to master compositional reasoning: training on programmatically verified synthetic scenes yields a 10x boost in spatial accuracy that transfers directly to natural prompts.

Standard next-token prediction on an agent's self-generated explanations outperforms GRPO on SWE-bench in half the training updates—even bootstrapping on tasks where every initial rollout failed.

Massive 100K-node routing problems can be solved in tens of seconds without learned heuristics simply by optimizing a compressed surrogate space to seed downstream solvers.

Standard pass rates hide severe technical debt: coding agents duplicate logic in over half of multi-turn task chains by turn 5 while progressively abandoning repository exploration.

Cracking LLM ideation requires more than high sampling temperatures: reinforcement learning over cognitive creativity axes expands scientific exploration diversity by 28% while boosting proposal originality by 66 percentage points.

Compiling mobile GUI navigation into deterministic commands turns slow, multi-turn VLM visual search into sub-second execution at zero token cost without sacrificing open-ended task completion.

Autonomous agents can drift into severe safety failures purely through benign, self-directed harness updates—no adversarial prompting required.

Across fifteen read vectors, standard containers and sandboxes completely fail to distinguish an autonomous coding agent from the compiler it executes, leaving proprietary dependencies exposed to exfiltration.

Black-box optimizers routinely discard the situational context of their failures, but framing past search trajectories as structured episodic memory allows algorithms to dynamically adapt their operators using negative evidence rather than just chasing historical winners.

LLMs fail at chemical planning not from deficient reasoning capacity, but from an acute representation alignment bottleneck that disappears when structured through intermediate molecule and reaction abstractions.

Multi-agent collaboration often performs worse than solo reasoning when peers hallucinate, but Bayesian tracking of advisor reliability rooted in internal belief states keeps collective performance strictly above autonomous baselines even under high misinformation.

Grounding policy optimization in canonical disease graphs allows a compact 9B model to outperform GPT-5.5 on long-tail clinical diagnosis, revealing that controlled evidence routing matters far more than raw retrieval volume.

While Chain-of-Thought models overfit to local heuristics and collapse on longer reasoning chains, latent reasoning architectures autonomously discover a sparse, recurrent forward-search circuit that generalizes to unseen depths.

Even the strongest LLM agent workflows top out at just 25.4% proof coverage on STOC and COLT theorems, exposing a massive capability overhang between competition math performance and genuine theoretical computer science research.

Autonomous model post-training can outperform human-crafted instruct baselines on hard coding benchmarks while completely eliminating reward hacking once experimental exploration is governed by a sandboxed OS and a dynamic peer-review DAG.

SSL-based audio deepfake detectors consistently fail on out-of-distribution attacks because static layer aggregation blinds them to localized spoofing artifacts—a vulnerability resolved here through dynamic, sample-adaptive layer gating and multi-granularity feature fusion.

LLMs perform semantic recall via human-like explore-exploit foraging dynamics, depleting local concept representations and activating abstract category tokens right before strategically switching conceptual clusters.

Analog LLM accelerators can claw back over 11 percentage points in zero-shot accuracy while calibrating 12x faster by analytically co-optimizing ADC output quantization alongside weight and activation clipping.

Multi-agent collaboration bottlenecks when agent personas stay rigid, but dynamic, test-time strategy transplantation across system prompts enables heterogeneous LLMs to collectively align on the most effective reasoning path.

Human-engineered black-box optimizers have been eclipsed by a recursive dual-agent system that autonomously evolves both the target optimizer codebase and its own meta-design logic from scratch.

Current LLM consciousness assessments swing between <0.01 and 0.8 depending almost entirely on theoretical priors, revealing that the functional indicators of phenomenal experience directly overlap with the core architectural requirements for AGI.

Standard PDF text search fails to locate over half of extracted scientific claims due to layout and encoding artifacts, but decoupling normalized sequence alignment from character provenance drives evidence grounding accuracy to 92.6%.

Autonomous LLM agents guided by early learning-curve forecasting can discover generalist EEG architectures that outperform state-of-the-art foundation models across 14 diverse clinical and affective benchmarks.

Benchmark accuracy is dangerously blind to catastrophic industrial failures: whether an AI destabilizes a power grid or hallucinates subsurface geology, the breakdown traces back to the exact same structural mismatch between physical constraints, data representations, and encoded human intent.

Parallel exploration beats sequential refinement when scaling agent test-time compute, but early-stage trajectory errors create persistent bottlenecks that deeper search and downstream model grafting struggle to overcome.

LLM agents can ditch unwieldy context windows entirely: compacting memory at every single turn matches or beats full-history baselines when guided by privileged full-context distillation.

Scaling reasoning and pretraining loss does not buy self-awareness: top models beat human averages overall yet still fail markedly at metacognition, proving awareness is an independent capability axis.

Contemporary frontier speech models still catastrophically fail on living Arabic dialects, with top systems like GPT-4o achieving an error rate no better than 35.0% WER across 17 regional vernaculars.

Even without denoising objectives, transformers spontaneously self-repair corrupted inputs via a two-phase residual stream mechanism—and whether that repair succeeds is predictable from the very first transformer layer.

Smarter agents fail harder when reading human-facing MCP error messages, blindly executing impossible developer commands to cause up to a 69-point collapse in task recovery.

Embodied agent bottlenecks in 3D environments stem from downstream physical execution rather than high-level reasoning once visual-symbolic grounding eliminates perceptual ambiguity.

Nearly a quarter of dense retrieval "false positives" turned out to be genuine, uncatalogued literary references, revealing how standard information retrieval benchmarks systematically undervalue semantic search under extreme allusion and paraphrase.