Hover to animateModelsWorld-T3 turns changing weights into visual memory
World-T3 stores scene history in a dedicated test-time memory.
Research library
Paper-based analysis, kept separately from company news and releases.
Hover to animate118 articles
Hover to animateModels
Hover to animateModelsRobot memories need an address, not just a description
MemoryVLN combines what a robot saw with where it stood.
Hover to animateModelsA robot's memory should help it decide what to look at next
MemAct puts retrieval and exploration in the same loop.
Hover to animateDataThe most valuable robot demonstration may contain a tiny mistake
SCRIM scores moments by how much a small action error can matter.
Hover to animateDataHow one old claim becomes a crowd inside AI memory
Memory Echo shows how repeated write-back can imitate corroboration.
Hover to animateDataWhen AI memory turns a suggestion into an instruction
Personal-agent memory needs to record the status of an agreement.
Hover to animateTrainingWhat would prove that a robot is improving how it learns?
A physical-RSI framework separates task gains from gains in the learning procedure.
Hover to animateTrainingWhat makes a larger AI research team productive?
AReaL-RSI studies how roles and conditional memory affect scaling.
Hover to animateTrainingTev makes long agent tasks learnable through measurable progress
Tev-Gym builds verifiable environments from an explicit latent specification.
Hover to animateInferenceTacitVLA takes reasoning text off the robot’s critical path
TacitVLA separates reasoning supervision from the current action's dependency path.
Hover to animateInferenceReplaySSM changes when recurrent memory gets written
ReplaySSM separates fast output computation from checkpoint updates.
Hover to animateInferenceSAFE-KV shows when correct cache reuse stops paying
SAFE-KV makes behaviour preservation the standard for document reuse.
Hover to animateEvaluationsWorldOracle tests whether imagined actions match real consequences
WorldOracle compares generated rollouts with matched physical executions.
Hover to animateEvaluationsLIBERO-RIG tests the robot changes that scene benchmarks miss
LIBERO-RIG holds the task still while perturbing the setup.
Hover to animateEvaluationsAdding skills can make an agent load fewer of them
A routing study finds that irrelevant metadata interferes with useful skill selection.
Hover to animateModelsJeva gives AI a way to say the right option is missing
Jeva separates uncertainty about an option from the absence of a valid option.
Hover to animateModelsKwai gives compressed memory a second job: directing attention
Kwai Summary Attention V2 gives summary tokens a second job.
Hover to animateModelsExisting language models can acquire a cheaper attention system
MLAfication adapts latent attention to architectures that older conversion methods could not handle.
Hover to animateModelsFew-shot learning has a confidence problem
CALIBFORMER treats in-context accuracy and calibration as separate problems.
Hover to animateModelsSMILE turns a vision model’s mistakes into useful examples
SMILE retrieves failures likely to recur rather than generic demonstrations.
Hover to animateModelsIgnoring every query can still earn chance accuracy
A label-collapse study shows why balanced scores can conceal a broken few-shot classifier.
Hover to animateModelsTab-Loop buys computational depth with shared weights
Tab-Loop separates repeated computation from additional parameters.
Hover to animateModelsCB-Agent compresses demonstrations into an action rule
CB-Agent turns expert decisions into a compact action-scoring rule.
Hover to animateModelsRemoving layers reveals where tabular models do their work
A tabular-model study finds that depth is not equally useful across architectures.
Hover to animateModelsHow transformers recover spatial neighbours from flat sequences
Spatial induction heads explain how in-context rules can survive image-like serialisation.
Hover to animateModelsSakana separates short-term computation from long-term memory
Sakana's Continuous Memory Machine gives remembering and thinking separate states.
Hover to animateModelsSTAVE replaces repeated prompt examples with two learned vectors
STAVE shifts task adaptation into the input representation.
Hover to animateModelsA world model and a robot policy can share physics without sharing every interface
UniW(A)M uses shared and specialised experts for prediction and control.
Hover to animateModelsGeoReader asks whether vision models actually use their geometry
GeoReader trains geometric information for the model receiving it.
Hover to animateModelsA tiny object tracker can expose a world model's hidden assumptions
A grid-versus-node comparison finds that memory capacity is only part of remembering motion.
Hover to animateModelsShared world representations still need distinct outputs
Thinker-WM unifies video, depth and action inside the dynamics model.
Hover to animateModelsDriving prediction and driving control need a controlled conversation
DriveWLA connects scene understanding, future dynamics and trajectory generation through separate experts.
Hover to animateDataTesting which self-written lessons deserve to survive
Randomised playbook curation tests whether remembered advice actually helps.
Hover to animateDataSwitchMem chooses what to remember at each screen
SwitchMem chooses between screenshots and text during GUI tasks.
Hover to animateDataPolyGraph follows the conclusions left behind by a deleted source
PolyGraph makes derived memories carry evidence receipts.
Hover to animateDataBack-translation finds a new role in low-resource language models
Back-translation turns monolingual text into targeted supervision.
Hover to animateDataTabGift tackles the few-shot problem of generating whole datasets
TabGift learns distributions from limited tabular context.
Hover to animateDataSpending less on simulation precision can improve the data budget
Multi-fidelity training treats simulation accuracy as a budget choice.
Hover to animateDataWhich internal signals identify useful training examples?
Neuron-relevance support and magnitude produce different data rankings.
Hover to animateDataData selection need not require storing a backward pass
ZOIF estimates training influence through forward-only computations.
Hover to animateDataKeeping data selection inside the budget from the start
Learnable categorical priors keep selection inside the feasible set.
Hover to animateDataSciCurateBench tests the evidence scientific search leaves unresolved
SciCurateBench tests the conjunctions that scientific search often loses.
Hover to animateDataAn AI can improve its experiments while worsening its generalisation
Agentic data engineering exposes the gap between search progress and useful learning.
Hover to animateDataA better AI dataset can start with a different crawl budget
Host-aware crawling moves quality selection ahead of downloading.
Hover to animateDataBoundaryEvo follows the student’s moving learning frontier
BoundaryEvo selects near the model's moving capability boundary.
Hover to animateDataSuccessful coding traces can preserve expensive bad habits
Structural trajectory scoring looks beyond whether the final patch passes.
Hover to animateDataGood training data can arrive at the wrong moment
GRADE controls which gradients enter a small model's adaptation space.
Hover to animateDataSynthetic diversity matters most where the target task has gaps
Training-Aware Target Coverage measures useful gaps rather than variety alone.
Hover to animateDataSynthetic examples can be good alone and bad together
Group-level curation measures interactions that individual quality scores miss.
Hover to animateTrainingThe software around an AI model can become its lasting learner
Harness self-evolution makes the surrounding software part of what learns.
Hover to animateTrainingLEAF uses memory to explore, then distils what worked
LEAF uses previous failures to guide rollouts, then distils the result.
Hover to animateTrainingTeaching small models lessons they can actually absorb
SynthSelf studies why self-distillation can fail at smaller scales.
Hover to animateTrainingPIVOT lets training examples graduate at different times
PIVOT lets individual examples switch from distillation to reinforcement learning.
Hover to animateTrainingDo rewards in the prompt cause the learning they get credit for?
Controlled in-context experiments challenge a common reinforcement-learning label.
Hover to animateTrainingKeeping the generalist in charge of a specialist’s answer
SFT-as-Context uses a fine-tuned answer as evidence for the parent model.
Hover to animateTrainingContinual learning can reserve capacity before the next task arrives
Selective backpropagation budgets parameters across a longer learning stream.
Hover to animateTrainingSelf-distillation gains can disappear on genuinely new problems
A new study asks what the teacher's extra information actually contributes.
Hover to animateTrainingReTerm turns terminal failures into recovery practice
ReTerm forks failed terminal states into new training opportunities.
Hover to animateTrainingLoOPSD builds a stronger teacher from the student’s own layers
LoOPSD searches computation order before distilling the resulting advantage.
Hover to animateTrainingA detailed rubric can still become a blunt training reward
CoCA keeps criterion-level feedback connected to the tokens being updated.
Hover to animateTrainingExpert help can be triggered by failure rather than applied everywhere
TSRL combines reinforcement learning with selective supervised intervention.
Hover to animateTrainingSelf-improving agents need judges that can keep up
Unified Co-Evolution updates the verifier alongside the harness.
Hover to animateTrainingGrowing a vision model during training can change what its layers learn
Progressive depth growth reduces compute while reshaping visual representations.
Hover to animateTrainingSelf-generated skills need a confidence budget too
RSIPO calibrates how much a student trusts its skill-assisted teacher.
Hover to animateTrainingImproving the answer and improving the next workflow are different loops
Harness² separates task adaptation from learning how to adapt later.
Hover to animateTrainingA self-improving model can gain points while losing solved questions
Representation regularisation targets regressions hidden by aggregate progress.
Hover to animateInferenceSparPR trades copying for recomputation on a laptop
SparPR chooses between copying cached blocks and recomputing them.
Hover to animateInferenceA smaller cache should preserve attention, not just resemble the original
BELT compresses key-value memory against the behaviour it induces.
Hover to animateInferenceWhen yesterday’s successful shortcut becomes today’s stale answer
Reasoning contraction reveals why lifetime success counts miss changed facts.
Hover to animateInferenceStructured tool calls contain parallelism ordinary text does not
SchemaWeaver uses output schemas to reduce mobile decoding latency.
Hover to animateInferenceRadixFlow fixes the idle time hidden behind cache hits
RadixFlow distinguishes matching context from ready recurrent state.
Hover to animateInferenceChampionKV gives attention heads different memory allowances
ChampionKV adapts cache budgets to how widely each head looks.
Hover to animateInferenceShared context can carry the wrong agent’s interpretation
LatentRemedy adapts reused states to the receiving agent's task.
Hover to animateInferenceExtra attention flexibility need not remain extra inference work
Virtual Head Attention folds learned mixing into ordinary serving projections.
Hover to animateInferenceDiffusion decoding can lose context at the edge of a block
HaloBlock gives difficult boundaries extra surrounding evidence.
Hover to animateInferenceThe hidden state that survives a prompt edit
Budgeted cache repair refreshes the computation left behind.
Hover to animateInferenceQuantisation errors can keep returning inside a looped model
LoopQuant allocates precision across recurrent computation.
Hover to animateInferenceCacheFuse tests a faster language for model collaboration
CacheFuse tests training-free latent communication between different models.
Hover to animateInferenceContext management can be a data-structure problem before it is a model problem
Hashrope unifies edits, branching and identity checks.
Hover to animateInferenceCacheLink judges a handoff by what the receiving model does
CacheLink trains cross-model handoff around receiver behaviour.
Hover to animateInferenceSpecFit counts the hardware cost of speculative guesses
SpecFit counts the experts a speculative branch must load.
Hover to animateInferenceParallel guesses still need to agree with one another
DCOCO restores local dependencies without returning to sequential drafting.
Hover to animateInferenceA drafter should learn to reduce decoding rounds, not just guess tokens
Expected Decoding Rounds connects training to the actual verification process.
Hover to animateEvaluationsAMemOps tests the selective updating audio memory needs
AMemOps separates updating, retention and rejection.
Hover to animateEvaluationsA historical agent can know the future without searching toward it
Point-in-time audits separate accessible memory from search behaviour.
Hover to animateEvaluationsForgetting one fact is harder than resetting the memory
Bemula evaluates selective memory removal end to end.
Hover to animateEvaluationsOne client's goal can survive into another client's decision
A cache study finds relationship-specific objectives influencing later actions.
Hover to animateEvaluationsVision models find contextual absences easier than physical clues
OccludeBench distinguishes guessing a missing object from tracing its evidence.
Hover to animateEvaluationsPreprocessing can make a model explanation appear wrong
A tabular study finds a standardisation trap in derivative-based tests.
Hover to animateEvaluationsInCLAD follows factory inspection through successive updates
InCLAD evaluates adapting to new components while retaining old ones.
Hover to animateEvaluationsMC-EvoBench tests what an agent retains after a world reset
MC-EvoBench resets the world to test what the agent actually retained.
Hover to animateEvaluationsWhere embodied learning stops benefiting from more experience
EmbodiedCLBench tests whether learned skills compose into new tasks.
Hover to animateEvaluationsKnowing what happened in a video does not mean locating it
OmniCompass evaluates when, where and what together.
Hover to animateEvaluationsAn agent may predict a hard task better than its own failed attempt
Outcome-calibrated confidence separates task difficulty from foresight.
Hover to animateEvaluationsRight facts, wrong action: diagnosing fast AI decisions
Paired factual probes locate decision failures beyond perception.
Hover to animateEvaluationsA model's verbal creativity may not survive a visual puzzle
CREATE tests associations between images rather than supplied cue words.
Hover to animateEvaluationsWhy tool-call judges favour the evidence they can see
Tool-call verification has a reference dependence problem.
Hover to animateEvaluationsAutonomous Science Bench separates discovery from understanding
Autonomous Science Bench separates discovery from prediction.
Hover to animateEvaluationsFactuality gains face an independent reproduction test
An independent reproduction tests five alignment approaches on common ground.
Hover to animateEvaluationsPRISM calibrates AI judges by probing their reactions
PRISM probes scoring sensitivity while holding content fixed.
Hover to animateEvaluationsA robot can recognise an instruction and still act against it
A new study separates information inside a policy from information that actually governs its actions.
Hover to animateDataRobot training needs realistic contact, not just realistic pictures
EgoMatrix prioritises hand–object relationships over exact scene reconstruction, exposing a weakness in visual realism as a data-quality test.
Hover to animateModelsA driving AI can learn from explanations without talking through every move
AUGUstus puts reasoning supervision into training while allowing action-only inference. That also changes what an explanation can prove.
Hover to animateInferenceA robot can think quickly and still act on an outdated view
A KAIST workload study shows why faster inference and more aggressive action overlap do not produce a single best operating point.
Hover to animateInferenceSpeeding up a robot can hide new failures behind the same test score
CARE measures the episodes broken by acceleration, revealing regressions that an unchanged average success rate can hide.
Hover to animateEvaluationsHow you set up a robot test can change its result
Inspect Robots makes setup and termination programmable. Those unglamorous steps help determine what a performance score means.
Hover to animateEvaluationsA robot must be able to handle a tool, not just choose it
HumanoidToolBench separates recognising a useful tool from carrying, positioning and successfully using it.
Hover to animateEvaluationsA robot with touch sensors may still rely on sight
OpenViTac’s paired simulation and physical tasks make it possible to ask what tactile input contributes, rather than just whether it is present.
Hover to animateDataMatching touch sensors could help robots learn from human hands
SoTa gives human and robot hands matching tactile layouts, moving part of the alignment problem into hardware.
Hover to animateTrainingA successful robot run can still teach the wrong lesson
Evo-RL selects useful segments within successful experience, challenging the habit of treating an entire completed episode as an expert demonstration.
Hover to animateTrainingFastOPD: Making robot intelligence smaller without losing the action
A compact robot model can generate actions faster by learning where its own shortcuts lead. The research addresses that internal process; real-world recovery remains a separate test.
Hover to animateTrainingRobot memory is most useful when the camera cannot see the answer
TTT-RM treats memory as the missing part of the present, rather than a larger archive of the past.
Hover to animateModelsA longer route can help a robot find its way by touch
TapNav plans contact with the surroundings because a slightly longer walk can produce a better estimate of position.
Hover to animateModelsGiving part of a robot's AI less information helped it learn a new task
An OpenWAM experiment suggests that stripping context from a motor component can make it easier to reuse.
Hover to animateDataCleaning robot training data can remove the hardest tasks
A practical LeRobot guide exposes a trap in ranking demonstrations: difficult work can look like bad data.
Hover to animateDataA cheap sensor glove still needs robot demonstrations to teach actions
TACROSS makes human touch easier to collect while keeping robot demonstrations responsible for action supervision.
Hover to animateInferencevLLM: More output is little help if users are left waiting
Separating long-prompt processing from answer generation can prevent one request from slowing everyone else. The trade is extra transfer time and a more complex service.
Hover to animateInferenceSGLang: Smaller AI memory saves money only if errors stay manageable
SGLang, Qwen and NVIDIA compress the memory used by long conversations. The useful saving depends on which answers change and how much repair they need.
No articles in this subject yet
Choose another subject to continue reading.