PaperlayerView exact revision on arXiv

Source-linked AI summary

State of Thought Enables Endogenous Reasoning

Zhiren Gong, Yikun Hou, Zihao Zeng, Ming Xiao, Chau Yuen, Wei Yang Bryan Lim

arXiv:2609.16055v2cs.CLcs.AI

TL;DR

Existing test-time reasoning relies on externally imposed programs or search and does not use a model’s evolving internal state to organize evidence. SoT uses a compact state to govern evidence selection and stopping. It reports accuracy gains across LLM and VLM evaluations, alongside lower token use and latency in the cited VLM comparison.

  • Problem

    Test-time reasoning commonly relies on externally imposed reasoning programs or expanded search, while leaving the model’s evolving internal state unused to organize decision-relevant evidence.

    Source: §1 · p. 2

  • Method

    SoT reads a compact dynamics-geometric state from internal information transfer and uses a lightweight controller on a frozen backbone to select historical evidence for subsequent reasoning.

    Source: §1 · p. 3, §2.2 · p. 3

  • Results

    Across 16 LLM datasets SoT improves accuracy over mean baselines; on VLMs it gains 3.8 points over reasoning baselines, using 74.9% fewer completion tokens and 73.5% lower latency than search-based methods.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The reported results support state-conditioned evidence organization as a route to reasoning improvement across text and multimodal evaluations and limited-access settings.

    Source: §1 · p. 3

    SoT’s compact four-dimensional state may under-specify unusually long, highly branching, or richly grounded trajectories.

    Source: §S.1 · p. 49

Abstract

from arXiv · show

Test-time compute has emerged as a major approach to improving the capabilities of Large Language Models (LLMs). However, existing test-time reasoning paradigms rely heavily on externally imposed control, either through fixed reasoning programs or through costly expansion in constrained search spaces, limiting both generalization and efficiency. We propose State of Thought (SoT), a new reasoning paradigm that enables endogenous reasoning in LLMs, with the model's internal reasoning state governing how reasoning unfolds. Concretely, SoT extracts a compact dynamics-geometric state from the model's internal information transfer and uses a 582-parameter controller on frozen backbones to selectively activate historical reasoning support useful under the current reasoning state, framing reasoning as a state-conditioned process over evidence rather than an externally prescribed token chain. Across quantitative (1.34x), general (1.62x), symbolic-and-code (1.76x), and long-context (2.51x) reasoning on 3 LLMs and 16 datasets, SoT consistently improves mean-baseline accuracy while reducing generated tokens by 62.6% and end-to-end latency by 44.6%. Across 2 VLM scales and 3 reasoning tasks, it improves mean accuracy by 3.8 points over reasoning baselines, with 74.9% fewer completion tokens and 73.5% lower latency than search-based methods. Under constrained access, SoT retains 38.2%/36.5% mean accuracy gains in training-free/embedding-only settings, while trajectory-only judging reaches 84.1% agreement across 3 API models. Together, endogenous state-driven reasoning provides a generalizable and efficient alternative.

1 Introduction

Existing test-time reasoning relies on external programs or search and does not use the model’s evolving internal state to organize evidence. SoT uses that state to select evidence and guide reasoning, with reported gains across text and multimodal evaluations.

Motivation

  • Existing reasoning approaches use externally prescribed programs or expanded search, leaving the model’s evolving internal state unused for deciding what evidence is needed.The paper identifies this as a limitation for generalization across heterogeneous reasoning structures and efficiency under constrained inference.Source: §1 · p. 2
  • Prior work uses internal signals descriptively or for early stopping, while context retrieval and latent-space reasoning do not use endogenous state to organize needed evidence.The paper frames endogenous state as a signal for organizing decision-relevant history.Source: §1 · p. 2

Approach

  • SoT reads a compact dynamics-geometric state from internal information transfer and uses it to select historical evidence for the next decision.The state-conditioned evidence loop is controlled by a 582-parameter controller.Source: §1 · p. 3

Results

  • SoT improves accuracy across 16-dataset LLM benchmarks and, on VLMs, raises mean accuracy by 3.8 points over reasoning baselines while using 74.9% fewer completion tokens and 73.5% lower latency than search-based methods.The LLM evaluation covers three backbones; SoT leads in five of six VLM model–task settings.Source: §1 · p. 3

2 State of Thought

SoT formulates reasoning as a policy-governed trajectory in which a compact endogenous state controls which prior evidence remains active and whether reasoning continues. Its state and resulting activation and stopping patterns vary with the current reasoning regime.

2.1 Problem Formulation

  • The formulation treats test-time reasoning as a policy-governed trajectory that organizes prior-step evidence, with SoT using endogenous state to govern evidence organization and continuation.The continuous state controls which explicit evidence is carried forward without replacing language-space reasoning with an opaque latent chain.Source: §2.1 · p. 3

2.2 Endogenous Reasoning State

  • SoT reads a four-dimensional dynamics-geometric state from internal information transfer, with coordinates for local structure, progress, directional consistency, and predictive uncertainty.The state-space comparison describes SoT as broader and more adaptive than externally structured strategies, which occupy distinguishable regions.Source: §2.2 · pp. 3–4, Figure 3 · p. 4

2.3 State-Conditioned Reasoning

  • State-conditioned evidence selection and stopping are joint outputs of SoT’s policy: one chooses active historical evidence, and the other decides when to terminate reasoning.Different endogenous state clusters correspond to distinct sparse activation profiles and stop probabilities.Source: §2.3 · p. 4, Figure 4 · p. 5
  • SoT learns only its evidence-organization and stopping operators while keeping the backbone frozen; the controller has 582 parameters.Supervision is constructed offline from reasoning trajectories and includes signals for useful historical evidence and stopping.Source: §2.3 · p. 5

3 Experiments

Across text benchmarks, SoT improves accuracy while reducing inference costs; ablations, VLM transfer, and a case study illustrate its components and broader scope.

3.1 Experimental Setup

  • Evaluation covers three frozen LLMs and 16 benchmarks across four reasoning domains, plus two VLM scales on three reasoning tasks, using task-standard accuracy, generated tokens, and latency.The LLM backbones are Llama-3.1-8B, Qwen2.5-14B, and Mixtral-8x7B; the VLMs are Qwen2.5-VL-7B and 32B.Source: §3.1 · pp. 5–6

3.2 Main Results

  • SoT leads all four Llama-3.1-8B reasoning-domain averages, improving over each domain’s strongest non-SoT baseline by +10.3/+15.9/+6.8/+10.1 points and ranking best or tied-best on 13 of 16 datasets.Its average domain-level gain over the strongest competing methods is 10.8 points.Source: Table 1 · p. 6, §3.2 · pp. 6–7

3.3 Efficiency and Trade-off

  • Averaged across four domains, SoT uses 223.9 tokens and 9.8s latency versus 472.7 tokens and 29.0s for sampling/search, reducing tokens by 52.6% and latency by 66.4%.Figure 5 further places SoT on a higher-accuracy, lower-cost frontier than external strategies and sampling/search baselines.Source: Figure 5 · p. 6, Table 2 · p. 7, §3.3 · pp. 7–8

3.4 Ablation Studies

  • Removing evidence organization S sharply lowers QS and LCR accuracy, while removing stopping T raises S&C accuracy from 54.5 to 57.1 but increases tokens from 189.8 to 2640.2 and latency from 9.8s to 112.0s.Geometry and dynamics also contribute complementary signals, and the ablations attribute SoT’s strength to component synergy rather than any single component.Source: Table 3 · p. 8, §3.4 · p. 8

3.5 Transfer to Vision-Language Reasoning

  • Across six VLM model–task settings, SoT averages 87.5% accuracy and leads in five, while reducing end-to-end latency by 73.5% relative to SC and BoN.At 32B it exceeds the strongest alternative by 0.5–0.7 points on all three tasks while using 3.1–4.6× fewer tokens than search-based methods.Source: Table 4 · p. 8, §3.5 · p. 8

3.6 Case study

  • The case study shows SoT selectively retaining evidence still needed for the next decision, suppressing obsolete support, and raising stop readiness as constraints are resolved.The example illustrates a closed loop in which state organizes support, support shapes the next step, and that step updates the state.Source: Figure 7 · p. 9, §3.6 · p. 9

4 Exploratory Extensions of State of Thought

Limited-access variants test whether SoT depends on learned control or hidden-state access, while SOT-JUDGE evaluates correctness from completed reasoning traces.

4.1 SoT under Limited Access

  • Training-free SoT and SoT-Embed remain competitive, with macro-averaged accuracy approximately 17.5% and 17.9% below full SoT, respectively.Both variants are comparable to or slightly better than the Top-3 baselines on many Llama-table tasks; the variants use fixed state-derived rules or text embeddings in the control loop.Source: §4.1 · p. 9, Table 5 · p. 10

4.2 SoT as a Trajectory-Level Judge

  • SOT-JUDGE improves average agreement over the strongest baseline by +17.7, +24.1, and +18.5 points on GPT-5.4, Claude Opus 4.7, and Qwen-Max, respectively.It predicts correctness from embedded sentence trajectories without extra rollouts, self-critique passes, or stronger external judge calls; it is weaker than standard judges on GU and S&C but reaches 81.0–99.0% agreement on LCR.Source: Table 6 · p. 10, §4.2 · p. 10

5 Conclusion

The conclusion presents SoT as state-conditioned evidence organization that governs both active support and reasoning termination; the appendix lays out its algorithmic, theoretical, and experimental support.

5 Conclusion

  • SoT shifts test-time reasoning control from external token programs to endogenous state-conditioned evidence organization, using evolving state to gate active evidence and determine when reasoning terminates.The conclusion frames this as a compact continuous state governing discrete, inspectable reasoning steps across text and multimodal benchmarks.Source: §5 · p. 10

Appendix Overview

  • The appendix provides algorithmic details, theoretical analysis, experimental setup, extended results, and discussion to make the controller reproducible and the reasoning claim evaluable.Its parts cover the inference loop and controller-facing state, theoretical arguments for state-conditioned control, evaluation protocols, and additional analyses.Source: Appendix Overview · p. 16
  • Appendix materials specify the SoT and trajectory-judge variants, formalize state-conditioned evidence organization and stopping, and collect datasets, baselines, hyperparameters, and qualitative case studies.They also include limitations and future work.Source: Appendix Overview · p. 16

A Implementation Details of State-Conditioned Reasoning

SoT uses a closed-loop controller to select historical evidence and decide whether to continue from the model’s current endogenous state. Its compact state summarizes sentence geometry, trajectory dynamics, and predictive uncertainty.

A Implementation Details of State-Conditioned Reasoning

  • The controller reads a four-dimensional state from each sentence-level reasoning unit and uses it to organize evidence and control continuation.The frozen backbone generates the sentence; the controller operates on the evolving sequence of sentence-state pairs.Source: A Implementation Details of… · p. 17
  • A state-conditioned stopping head applies a learned threshold to the current state, rather than following a fixed external depth schedule.The same state also conditions which historical evidence remains active for the next decision.Source: A Implementation Details of… · pp. 17–18
  • SoT alternates state-conditioned evidence selection, sentence-level generation, and stopping, making the reasoning trajectory path-dependent rather than a fixed script.The new reasoning unit changes the next state, while the active evidence changes what the model generates next.Source: A Implementation Details of… · pp. 17–18

B.1 State construction

  • The state combines within-sentence dispersion, movement between adjacent units, directional alignment across movements, and mean predictive entropy.Dispersion captures whether a unit is concentrated or diffuse; movement and alignment distinguish trajectory change from directional stability; entropy summarizes local predictive uncertainty.Source: §B.1 · pp. 18–19

B.2 Why four coordinates suffice for control

The four state coordinates provide distinct inputs for evidence selection and stopping. Interventions and state-matching tests connect these coordinates to controller behavior and task outcomes.

B.2 Why four coordinates suffice for control

  • Each coordinate captures a non-interchangeable control signal: dispersion versus predictive uncertainty, and movement magnitude versus directional consistency.The four observables provide a current-unit snapshot, trajectory history, and predictive-readiness readout for evidence organization and continuation decisions.Source: §B.2 · p. 19, Table 7 · p. 19
  • Zeroing δ_t during generation lowers accuracy by 6.7–22.5 points, while zeroing v_t makes trajectories 4–5× longer on Llama-3.1-8B and Qwen2.5-14B.Across 120 examples each from GSM8K, DROP, and MATH, dispersion was linked to answer quality and displacement to halting; every coordinate also changed a substantial fraction of selections.Source: §B.2 · pp. 19–20
  • Permuting historical states across candidates reduces accuracy from 79.0 to 62.4 on GSM8K, 38.6 to 24.6 on DROP, and 60.0 to 38.5 on MATH.The permutation preserves candidate text, position, and inference budget while breaking only the state–unit correspondence.Source: §B.2 · p. 20
  • The four-coordinate state supports relational evidence selection and direct stopping with 582 trainable parameters total, while the backbone remains frozen.The selection head has 577 parameters and the stopping head has 5.Source: §B.2 · p. 20

C Lightweight Training of State-Conditioned Operators

SoT fits lightweight evidence-selection and stopping operators offline from trajectories generated by a frozen backbone. Training uses state-paired historical evidence and replay-derived supervision, with evaluation examples excluded.

C Lightweight Training of State-Conditioned Operators

  • The offline pool consists of sentence-segmented rollouts from the frozen backbone, with state-normalization statistics estimated from the same trajectories.For VLM transfer, evaluation examples are excluded and fitting remains confined to the lightweight controller.Source: C Lightweight Training of… · p. 20
  • Offline replay supplies retention targets for useful state-compatible evidence and stopping targets based on accumulated support and trajectory maturity.On correct trajectories, stop mass appears when the trace is mature and locally confident; on incorrect trajectories, it is deferred until the trace is late and stagnating.Source: C Lightweight Training of… · pp. 20–21
  • SoT transfers beyond the trajectory-collection datasets, and its state-conditioned principle remains effective without learned controller fitting.Trajectory quality affects how sharply the controller can be supervised, but SoT does not learn to mimic a fixed reasoning trace.Source: C Lightweight Training of… · p. 21
  • SoT trains only its 582-parameter controller bundle, leaving backbone weights fixed and excluding evaluation examples and test-time rewards from fitting.Training uses 12 warm-up epochs followed by 400 refinement steps; selector AUC across five problem-held-out refits is 0.779 ± 0.005.Source: C Lightweight Training of… · pp. 20–21

D SoT under Limited Access and Limited Training

The paper examines SoT under limited access, post-hoc judging, and theoretical policy-class comparisons. These sections describe alternative control signals and state the assumptions under which state-conditioned control can outperform fixed policies.

D SoT under Limited Access and Limited Training

  • Training-free SoT retains the state interface but replaces learned evidence selection and stopping with fixed state-compatibility scoring and stable, low-uncertainty stopping.SoT-Embed instead derives control inputs from sentence embeddings and trajectory summaries when hidden states are unavailable.Source: D SoT under Limited Access and Limited… · p. 22

E SoT-Judge

  • SOT-JUDGE predicts reference correctness from sentence-embedded trajectories and pooled trajectory statistics, without hidden-state access or extra generator calls.The paper frames it as a lightweight black-box alternative to majority-vote selection or stronger-model judging.Source: E SoT-Judge · pp. 22–23

Part B: Theoretical Analysis

F Policy-Class Separation

  • The policy-class comparison concludes that changing prompts or enlarging search may remain within an external control class, while SoT changes the control variable to endogenous state.The argument does not claim every external method is weak or incapable of improvement.Source: F Policy-Class Separation · p. 25

G State-Conditioned Evidence Organization

  • Under the stated assumptions, state-conditioned evidence selection contains a Bayes-optimal selector, whereas any fixed evidence operator that ignores the current state is strictly suboptimal.The result assumes that evidence demands differ across regimes for a shared history and that the state is sufficient to select relevant evidence.Source: G State-Conditioned Evidence… · pp. 25–26

H Sparse Evidence Organization

The section argues that regime-conditioned reasoning should retain sparse, decision-relevant history rather than preserve the full history. Under its stated utility assumptions, removing interfering evidence while preserving helpful evidence reduces cost without reducing reward.

H Sparse Evidence Organization

  • The argument assumes history decomposes into helpful and interfering evidence, with monotone reward and interference terms and a positive cost for carrying context.The proposition fixes a realized history and reasoning regime; helpful evidence is preserved while at least one interfering unit is removed.Source: H Sparse Evidence Organization · pp. 26–27
  • Under Proposition H.1, a sparse operator preserves helpful evidence, removes at least one interfering unit, and strictly reduces the cost term without increasing interference.The proof concludes that the reward term is unchanged, so the sparse operator improves the stated utility over full-history retention.Source: H Sparse Evidence Organization · p. 27
  • State-aligned sparsity is presented as a way to jointly improve accuracy and efficiency, not merely as generic compression of the history.Full-history retention can preserve regime-mismatched interference alongside useful support.Source: H Sparse Evidence Organization · pp. 26–27

I Closed-Loop Reasoning Control

The section formalizes SoT as closed-loop control over evidence retention and stopping, with decisions conditioned on the realized endogenous state. It also lists the dataset categories used to organize the evaluation.

I Closed-Loop Reasoning Control

  • When different realized states require different actions, closed-loop control can attain strictly greater expected value than any fixed open-loop schedule.Every open-loop policy is included in the closed-loop class; strict improvement follows when optimal actions vary across a positive-probability set of states.Source: I Closed-Loop Reasoning Control · p. 28

J State-Driven Stopping

  • Stopping is part of the same state-conditioned control problem: continue when another step’s marginal value is positive and stop otherwise.A fixed depth or step-index-only rule is generally suboptimal when the continuation advantage changes sign across realized states.Source: J State-Driven Stopping · pp. 28–29

K.1 Category organization

  • The LLM evaluation groups datasets into quantitative, general-understanding, symbolic-and-code, and long-context reasoning categories.Multimodal transfer uses A-OKVQA, AI2D, and M3CoT for knowledge-intensive visual QA, science-diagram reasoning, and multi-domain multimodal reasoning, respectively.Source: §K.1 · p. 30

K.2 Metrics

This section defines the primary task metrics and the efficiency measures reported across models.

K.2 Metrics

  • The headline “accuracy” denotes each benchmark’s task-standard final metric, and Table 9 summarizes dataset organization, split size, and primary metric.The primary metrics are EM for classification-style and short-answer tasks, F1 for long-form QA and evidence aggregation, and P@1 for code generation.Source: §K.2 · pp. 30–31, Table 9 · p. 30
  • Efficiency is measured by average generated reasoning tokens per example and end-to-end wall-clock inference time under the same hardware and decoding.Latency is reported as the end-to-end inference time.Source: §K.2 · p. 31

L Model Configuration and Evaluation Protocol

The supplementary sections specify evaluation and baseline configurations, interpret state-conditioned control behavior, and report additional backbone results. Across those additional backbones, SoT leads in several domain averages and its reported advantage spans different model characteristics and task families.

L Model Configuration and Evaluation Protocol

  • Evaluation uses frozen Llama-3.1-8B, Qwen2.5-14B, and Mixtral-8×7B text backbones and frozen Qwen2.5-VL-7B/32B multimodal backbones, with shared splits, decoding, extraction, and scoring.Learned components, including SoT, are fit only on pre-specified training-side data, excluding evaluation questions.Source: L Model Configuration and Evaluation… · p. 31

M.1 Direct generation

  • Baselines cover direct greedy answering and externally scripted reasoning, including CoT, Plan-and-Solve, and Self-Refine.The comparison also includes breadth or search methods such as Self-Consistency, Best-of-N, constrained beam, and MCTS.Source: §M.1 · p. 31, §M.2 · p. 32, §M.4 · p. 32

M.3 Memory-oriented methods

  • Memory-oriented baselines test generic context reduction, while COCONUT and GRPO-SP test latent-space and optimization-based alternatives to inference-time state-conditioned control.The comparison logic positions these families as alternatives to evidence control, rather than treating their goals as identical.Source: §M.3 · p. 32, §M.4 · pp. 32–33

N Hyperparameter Configuration

  • Deployment-side score smoothing adjusts ranking signals before thresholding, and about 99.6% of trajectories terminate through the learned stop policy before the maximum reasoning-step cap.The hard cap is described as a safeguard, not the source of the reported efficiency gains.Source: N Hyperparameter Configuration · pp. 33–34

O.1 State-space trajectories across reasoning paradigms

  • Across paradigms, Figure 8 visualizes projected state trajectories; the interpretation reports distinguishable trajectory regions and broader, more adaptive state-space coverage for SoT.The analysis does not claim that each paradigm maps to a discrete symbolic state label.Source: Figure 8 · p. 35, §O.1 · p. 35

O.2 State clusters, sparse memory, and stopping

  • Figure 9 links clustered states to distinct evidence sparsity and stopping profiles, with cluster labels assigned after clustering rather than used in fitting or evaluation.Profiles are examined through activation density, retained-unit timing, and stop probability; the reported differences are not attributed to generic recency bias or fixed-depth schedules.Source: Figure 9 · p. 36, §O.2 · p. 36

O.3 Correct versus incorrect trajectories under internal and embedding-trajectory SoT

  • Figure 10 compares correct and incorrect trajectory distributions across four task categories under full-state SoT and SoT-Embed; the reported separation persists with embedding-derived features.The passage presents trajectory-state separability as a potential reliability-monitoring signal when white-box internals are unavailable.Source: Figure 10 · p. 37, §O.3 · p. 37

P.1 Additional backbone results

  • Qwen2.5-14B SoT leads QS (66.1), S&C (76.8), and GU (81.4), while Mixtral-8x7B SoT leads S&C (51.7), GU (72.6), and LCR (23.5).On Qwen2.5-14B, all five GU datasets reach the table’s highest scores; on Mixtral-8x7B, QS is near-best at 46.3 versus 46.4. The reported cross-backbone pattern supports transfer across architectures and dataset families.Source: §P.1 · p. 37, Table 12 · p. 38, Table 13 · p. 38

P.2 Extended efficiency and trade-off results

Extended analyses place SoT near the accuracy–efficiency frontier across backbones and probe its components, limited-access variants, and trajectory-pool sensitivity. Case studies illustrate selective evidence carry across reasoning tasks, while the paper positions SoT as a distinct control formulation and identifies scope boundaries.

P.2 Extended efficiency and trade-off results

  • Qwen LCR token average was 4901.5 for SC versus 254.0 for SoT; across backbones, SoT stays on or near the accuracy–efficiency Pareto frontier.Search-heavy methods also incur large token and latency overheads, especially on long-context tasks.Source: §P.2 · pp. 38, 43

P.3 Additional ablation results

  • Removing evidence organization sharply lowers Qwen accuracy (QS 66.1→30.0; S&C 76.8→28.4), while removing stopping can raise GU tokens from 133.2 to 4175.0.Threshold tuning is negligible on Qwen and Mixtral, while state-component removals typically trade shorter outputs for weaker accuracy.Source: §P.3 · p. 43, Table 20 · p. 43, Table 21 · p. 44

P.4 Limited-access variants and SoT-Judge

  • On Qwen2.5-14B, Training-free SoT reaches an LCR average of 26.2 versus 25.8 for Top-3 baselines.For trajectory judging, LCR agreement is 81.0 for GPT-5.4 versus ≤16.7 for baselines, 99.0 for Claude 4.7 versus ≤9.3, and 82.2 for Qwen-Max versus ≤14.8.Source: §P.4 · pp. 44–45, Table 22 · p. 44, Table 23 · p. 44

P.5 Sensitivity to trajectory quality

  • Weaker trajectory pools disproportionately hurt GU and LCR while inflating tokens and latency, whereas QS and S&C accuracy largely stay near plateau.Pool curation affects whether supervision yields a cheap, reliable trajectory-level control signal more than it uniformly rescales backbone capability.Source: §P.5 · p. 45, Table 25 · p. 45

Q.1 Case 1: Quantitative Reasoning (QS)

  • In QS, active support follows a staged numeric workflow; in GU, late steps retain locally decisive intermediates while dropping earlier scaffolding.The QS example raises stop probability once competing subtotals are fixed; the GU example avoids carrying verbal distractors in a full-history transcript.Source: §Q.1 · p. 46, §Q.2 · p. 46

Q.3 Case 3: Symbolic and Code (S&C)

  • In S&C, later steps reanchor on conclusions and skip obsolete rule instantiations; in LCR, sparse retrieval fuses clauses relevant to a contrastive why question.The visual-language case applies the same interface to chart QA through panel-wise extraction, numeric-gap reasoning, rising stop probability, and sparse dependencies.Source: §Q.3 · p. 46, §Q.4 · p. 47, §Q.5 · p. 47

R.5 Positioning summary

  • SoT differs from scripted reasoning, descriptive hidden-state analysis, and generic context compression by using the current endogenous regime to control evidence selection and continuation.It routes explicit sentence-level reasoning through a compact controller rather than replacing it with a fully latent process or merely shortening chains.Source: §R.1 · p. 48, §R.2 · p. 48, §R.3 · p. 48

S. Limitations and Future Work

  • The evaluation covers multiple text backbones and two scales of one multimodal backbone family; compact R^4 states may under-specify unusually long, branching, or richly grounded trajectories.Future directions include richer but still compact states, adaptive memory units, and broader multimodal and interactive settings.Source: S Limitations and Future Work · p. 49, §S.1 · p. 49, §S.2 · p. 49
Loading 2609.16055v2…