Amazon Science

Multimodal unlearning could revolutionize how we handle sensitive data in AI, enabling targeted removal without sacrificing model performance.

Thoughtful feature curation reveals that structural code complexity is a far stronger predictor of deployment risk than traditional change volume metrics.

Current MLLMs struggle with Bangla form comprehension, missing key granular details that could hinder their real-world application in low-resource languages.

Achieving over 42% recall in semantic video communication could redefine how we transmit meaning in bandwidth-limited networks.

DBPP slashes peak memory usage by up to 25 times during large container image pulls, preventing OOM failures on GPU nodes.

TrajLoc achieves unprecedented trajectory adherence and visual fidelity in multi-object motion control, outperforming existing methods by isolating object trajectories with Gaussian heatmaps.

A unified creativity benchmark reveals that LLMs exhibit a single creativity factor, yet top human creators still outperform them in creative tasks.

Bi-NAS boosts recommendation accuracy while transforming user explanations into clear, personalized insights that enhance trust and engagement.

AutoRestTest outperformed all competitors in the SBFT 2026 Tool Competition, revealing significant advancements in automated REST API testing.

Data referencing errors plague LLMs even in structured tasks, but a lightweight critic model can boost accuracy by up to 12%.

Zero-shot generation of 360 panoramas is now possible without the costly fine-tuning or optimization typically required, unlocking new creative potentials in image synthesis.

Homography-based pose estimates can outperform traditional methods in planar scenes, revealing a new pathway for robust camera pose recovery.

Matched reference regimes for prosody evaluation reveal that traditional methods over-flag deviations, leading to misinterpretations in speech AI assessments.

Synthesizing 48,000 interaction trajectories without human input enables a humanoid robot to learn complex loco-manipulation tasks effectively.

Humans' innate ability to follow instructions may mirror the mechanisms that allow LLMs to excel at zero-shot learning, revealing a shared cognitive architecture.

Translating C interpreters to safe Rust can be done with minimal human intervention while completely eliminating memory vulnerabilities.

Axon can automatically synthesize high-performance tensor programs, drastically simplifying the optimization process for AI accelerators.

ForceBand transforms human muscle signals into precise force data, enabling robots to learn manipulation tasks with unprecedented accuracy.

A robust multi-agent scaffold can unlock latent capabilities in fixed models, enabling a remarkable 67.4% issue resolution rate on SWE-bench Pro—outpacing the previous best by over 8 percentage points.

Advanced RAG methods like GraphRAG and Agentic RAG can reduce token usage by up to 53%, but they don't always enhance generation quality as expected.

SurfBind's innovative surface-centric approach outperforms traditional methods, revealing the critical role of molecular surface interactions in epitope prediction.

Bayesian Contextual Bandits not only outperformed traditional models in real-time sorter optimization but also achieved a 2.03% reward uplift, showcasing its potential for dynamic warehouse environments.

Offline RL can boost warehouse operational efficiency by nearly 23% while cutting throttling times, revealing a game-changing approach to throughput control.

Exact-match retrieval metrics can mislead assessments of policy utility, as retrieved clauses perform nearly as well as gold-standard ones in decision-making tasks.

Large-scale human motion data can not only train robot controllers but also optimize the physical designs of robot hands, achieving superior performance in real-world applications.

S-JEPA sets a new standard in speech representation learning by achieving top performance with fewer parameters and without the cumbersome offline re-clustering process.

Graphical PLC programs can now be verified accurately in under 70ms, eliminating previous vacuous results and enhancing reliability in industrial automation.

The largest-ever verification campaign for Rust's standard library reveals significant vulnerabilities in unsafe code, underscoring the need for robust static verification methods.

Bridging the intent-execution gap reveals that performance metrics alone can obscure significant behavioral differences among AI models in problem-solving contexts.

FoundCause outperforms traditional causal discovery methods by explicitly modeling latent confounders, achieving superior accuracy in a single inference pass.

Achieving a 91 μs root recomputation time, Fractional Verkle Trees could redefine efficiency in blockchain state management.

Direct token-level self-distillation can backfire, but Sibling-Guided Credit Distillation redefines credit assignment to enhance long-horizon tool-use without amplifying harmful behaviors.

Semantic progress in dialogue can be quantified effectively without relying on large models, achieving human-level agreement on information gain across turns.

High-performance ML models can be reproduced with minimal information, revealing that they thrive in low-complexity regions and defy traditional overfitting concerns.

GLACIER achieves high predictive accuracy while drastically cutting down the computational costs typically associated with multimodal molecular property prediction.

Generated translation references outperform ground truth by 8.65 CEA100 points, showcasing a novel approach to literary translation challenges.

Sustained self-improvement in LLM agents is achievable through a novel adaptive framework that outperforms traditional methods in dynamic task environments.

Unlock hidden dynamics in noisy X-ray experiments: a fully convolutional autoencoder now efficiently denoises variable-sized correlation functions, even under photon-limited conditions.

LLMs can autoformalize specs well enough to pass standard tests, but still fail on subtle edge cases 26% of the time, a risk missed by LLM-as-judge evaluations.

ESBMC's journey from a research prototype to an autonomous verification kernel integrated with LLMs and deployed industrially at Lockheed Martin signals a paradigm shift towards AI-driven formal verification.

Multi-objective prompt optimization for LLM judges suffers from a staggering 59% drop in task-focus when gradients are combined, revealing critical design constraints.

LLMs can resolve merge conflicts nearly as well as Google's best, but still fail in over 40% of cases, revealing a surprising bottleneck in automating software development.

Open-source QUEST agents, trained solely on 8K synthetic tasks, rival or surpass proprietary research agents, proving that scaling data synthesis can unlock frontier performance.

LLMs can now automatically slim down and future-proof mathematical proofs, achieving 70% compression and 60% faster compilation by strategically rewriting them.

Semantic watermarks, embedded via AMR, survive paraphrasing attacks that obliterate token-level watermarks.

Turns out, the best template for documenting architectural decisions depends on whether you value conciseness (Nygard) or structural detail (MADR).

Multimodal models can now achieve state-of-the-art performance in real-world tasks like document understanding and audio-video comprehension with significantly reduced inference latency thanks to novel token-reduction techniques.

Directly embedding quantile tokens into input sequences leads to sharper and more accurate distribution predictions, outperforming traditional methods by a substantial margin.

Upcycling MoE models can achieve the same performance as larger fixed-size models while cutting GPU costs by 32%.

Targeted neuro-symbolic integration can reduce content bias in syllogistic reasoning, achieving over 94% accuracy while cutting content effects by 16%.

Current red-teaming efforts miss the forest for the trees: ARES reveals that safety failures often stem from a systemic breakdown between the LLM *and* the reward model, not just the LLM itself.

LLMs can now autonomously translate entire C projects to Rust with near-perfect accuracy, thanks to a novel agentic framework that dynamically navigates dependencies and iteratively verifies translations.

RAG systems are stuck in a factual echo chamber, ignoring the rich tapestry of opinions that shape real-world understanding.

Domain-specific fine-tuning can induce "agentic collapse" in LLMs, but a surprisingly small amount of agentic data from *another* domain can bring those general tool-use skills roaring back.

Forget wrestling with language-specific tooling: ReCodeAgent autonomously translates and validates entire code repositories across diverse languages with a 60% boost in test pass rates.

Speculative decoding's speed boost just got a whole lot bigger: DIVERSED dynamically loosens the verification constraints, letting more good tokens through and accelerating inference.

Prime Video's new anomaly detection system spots real incident-related services missed by traditional load testing, proving that synthetic traffic can't always predict live event behavior.

LLMs editing code are far more reliable and efficient when manipulating ASTs instead of raw text, slashing invalid patches and token costs.

Ditch the computational bloat: DeltaWorld slashes parameters by 35x and FLOPs by 2000x while generating more realistic video futures.

LLMs aren't culture-aware reasoners, but biased translators: they generate stereotyped metaphors and default to Western perspectives even when prompted with specific cultural identities.

LLMs can boost code performance by 25%, but only when working *with* compilers in a carefully orchestrated multi-agent system.

LLMs can automatically generate web vulnerability detection rules with surprisingly high accuracy, but only with careful validation and human oversight to mitigate overconfidence.

Recommending popular items isn't always what users want: SPREE steers sequential models to align with individual users' preferences for popular or niche content, improving recommendations.

Memory-augmented LLMs get a strategic upgrade: MemMA uses multi-agent reasoning to proactively guide memory construction and repair, leading to significant performance gains.

LLM-generated survey responses can be statistically accurate yet still miss the option most preferred by humans, highlighting a critical flaw in current evaluation methods.

Agentic LLMs are surprisingly vulnerable: a new framework finds successful attacks in 84% of attempts by escalating prompt injection techniques across multiple stages.

Achieve minute-level navigable video world models by combining the strengths of explicit 3D patch memory with implicit generative modeling.

Achieve near-full light throughput in spectral imaging with a novel oscillating dispersion technique and deep unfolding network, enabling high-fidelity reconstruction even under light-starved conditions.

Achieve 50% bitrate savings in ultra-low-bitrate image compression by cleverly turning image decoding into a next-frame prediction problem using video diffusion priors.

LoRA fine-tuning can significantly boost the voice cloning capabilities of LLM-based TTS systems, but only if the training data is acoustically diverse enough.

LLM reasoning research is inadvertently paving a dangerous path towards AI situational awareness and strategic deception, demanding a re-evaluation of current safety measures.

Recursive self-improvement can boost performance by 18% in code and 17% in reasoning, but only if you can keep it from going off the rails – SAHOO provides the guardrails.

MC3D models can now generalize to unseen camera configurations thanks to a new framework that explicitly accounts for spatial prior discrepancies.

Save 20% on LLM costs with <2% accuracy drop by strategically cascading a small model with a large one, guided by a confidence-calibrated SLM.

LLM-based recommender systems can trigger users' personal trauma, phobias, or self-harm history, but a new framework cuts these safety violations by 96.5% while maintaining recommendation quality.

LLMs can ace math problems while reasoning like a drunk toddler, with 82% of correct answers arising from unstable, inconsistent logic.

Safety classifiers for LLMs can catastrophically fail with even minuscule embedding drift, creating dangerous blind spots in deployed safety architectures.

Injecting knowledge graphs into LLMs boosts medical question generation by 8%, suggesting a simple way to patch up LLM knowledge gaps.

Despite matching or exceeding human expert performance on generating potential diagnoses, current MLLMs struggle to synthesize multimodal clinical evidence for final diagnosis, revealing a critical gap in their clinical reasoning abilities.

Forget Bonferroni: a new sequential testing approach slashes audit times for multi-stream ML systems, especially when anomalies are widespread.

Latent reasoning models often take shortcuts to achieve high accuracy, and stronger supervision, while mitigating this, paradoxically restricts the diversity of their latent representations.

Soft pseudo-labels, theoretically equivalent to hard labels when perfectly calibrated, tank performance in cross-domain semantic segmentation, motivating a new calibration framework.

Stop training your M3OD models on the same old entangled data: this method decomposes and recomposes objects, scenes, and camera poses to generate diverse training examples on the fly, boosting performance without needing more real-world data.

Forget fine-tuning: inject targeted time-series insights into general LLMs and watch their reasoning skills skyrocket by up to 26%.

Static benchmarks can be fooled by fluent text and aligned citations, but DREAM leverages agentic evaluation to expose the critical capability mismatch in assessing temporal validity and factual correctness of research agents.

LLMs may ace the test, but their uncertainty estimates are far from perfect, raising serious concerns about their reliability in high-stakes educational assessments.

An end-to-end system extracts funny scenes from movies with 87% accuracy, opening new avenues for automated content repurposing.

Give new e-commerce products a warm start by borrowing behavioral signals from their substitutes, boosting search relevance and product discovery.

Stop hand-rolling your multi-task learning to rank models: DeepMTL2R provides a ready-to-use framework with 21 SOTA algorithms and Pareto-optimal optimization.

Object hallucination in MLLMs can be significantly reduced by simply masking salient visual features during contrastive decoding.

MLLMs can now reason about road traffic accidents by fusing remote sensing imagery and structured data, unlocking interpretable insights previously inaccessible to traditional methods.

Pinpointing the root causes of supply chain anomalies just got easier: a Shapley value-based attribution mechanism rapidly decomposes simulation outputs into individual input effects.

Open-source LLMs can now autonomously optimize AI accelerator kernels, matching the performance of proprietary models at a fraction of the cost.

AI-generated feedback on student portfolios from GPT-4o and Claude-Sonnet-4 shows promise for high-stakes clinical assessments, but careful evaluation is needed to ensure accuracy and educational value.

LLMs evaluating job candidates exhibit significant bias against hedging language, docking candidates by 25.6% on average, even when the content is equivalent.

Achieve up to 39.6% FLOP reduction in LLM inference without retraining or architectural changes using QuickSilver's dynamic token-level optimizations.

By focusing on the most challenging examples, CRPO significantly boosts machine translation accuracy and data efficiency compared to standard preference optimization techniques.

A flexible headline generation model that boosts click-through rates by over 5% while allowing precise control over headline length.