Autonomous agents operating in complex, multi-agent environments must reason about what is true from multiple perspectives. Existing approaches often struggle to integrate the reasoning of different agents, at different times, and in different contexts, typically handling these dimensions in separate, specialized modules. This fragmentation leads to a brittle and incomplete reasoning process, particularly when agents must understand the beliefs of others (Theory of Mind). We introduce the Observer-Situation Lattice (OSL), a unified mathematical structure that provides a single, coherent semantic space for perspective-aware cognition. OSL is a finite complete lattice where each element represents a unique observer-situation pair, allowing for a principled and scalable approach to belief management. We present two key algorithms that operate on this lattice: (i) Relativized Belief Propagation, an incremental update algorithm that efficiently propagates new information, and (ii) Minimal Contradiction Decomposition, a graph-based procedure that identifies and isolates contradiction components. We prove the theoretical soundness of our framework and demonstrate its practical utility through a series of benchmarks, including classic Theory of Mind tasks and a comparison with established paradigms such as assumption-based truth maintenance systems. Our results show that OSL provides a computationally efficient and expressive foundation for building robust, perspective-aware autonomous agents.
Figures & tables
Figure 1. OSL-based agent architecture. Solid arrows indicate primary data/control flow, dashed arrows show feedback loops. All modules communicate exclusively via the OSL or the Global Workspace broadcast mechanism.
Figure 2. OSL cognitive cycle with nominal 100ms period. Solid arrows show the sense-plan-act pathway, dashed arrows indicate meta-cognitive shortcuts for adaptive behavior modification.
Figure 3. Runtime and memory comparison showing OSL’s competitive performance against established truth maintenance systems. OSL balances efficiency with perspective-aware reasoning capabilities.
System
Runtime (ms)
Mem. (MB)
Proc. Rate
Elem.
Conv.
Capabilities
OSL
6.87 ± 0.22
0.06
3494 ops/s
24
100%
Perspective-aware
ATMS
11.65 ± 0.81
0.36
2060 ops/s
154
100%
Assumption-based
DTMS
5.34 ± 0.37
0.09
4494 ops/s
154
100%
Dependency tracking
MEPK
67.72 ± 9.66
0.42
354 ops/s
120
100%
Probabilistic
Table 1. Performance comparison across baseline systems (24-element lattices, 5 trials).
Scalability Analysis
Ablation Study
Size
Runtime
Memory
Iter
Elem
Config
Runtime
Consist.
Cov.
Bel.
4
0.96ms
0.01MB
20
3.6
Full
7.60ms
-5.49
1.0
142
8
1.62ms
0.00MB
20
7.4
No MCC
5.06ms
-5.49
1.0
142
16
3.90ms
0.04MB
20
15.4
Limited RBP
4.39ms
-5.49
1.0
142
32
13.19ms
0.05MB
20
32.0
No Propagation
1.01ms
-6.89
0.0
72
64
34.82ms
0.17MB
19.4
64.0
Minimal
0.001ms
-6.89
0.0
72
Table 2. Scalability results and ablation study across configs.
Scenario
OSL result
Expected
Confidence
Sally–Anne (basic)
PASS
PASS
1.000
Sally–Anne with distractor
PASS
PASS
1.000
Nested belief (Level 2)
PASS
PASS
0.950
Multiple objects
PASS
PASS
1.000
Temporal belief change
PASS
PASS
0.975
False photograph
PASS
PASS
1.000
Table 3. Theory-of-mind test results across scenarios.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Lattice
Observers
Situations
Balanced
Avg Time
Std Dev
Affected
Max
Trials
Size
( μ s)
( μ s)
Elements
Affected
9
3
3
No
8.31
7.07
2.5 ± 2.3
8
30
16
4
4
No
9.79
7.84
3.5 ± 3.1
11
30
30
5
6
No
21.21
19.01
7.1 ± 6.9
23
30
56
7
8
No
13.91
13.57
4.8 ± 5.4
17
30
100
10
10
Yes
9.25
9.53
3.2 ± 3.7
19
30
Appendix
Table 4. RBP scaling performance results across all tested lattice sizes. “Avg Time” and “Std Dev” are computed over the 30 trials for each lattice configuration (units: μ s). “Affected Elements” reports the mean ± standard deviation of the number of lattice elements updated by RBP in a trial, and “Max Affected” reports the maximum updated count observed across the trials.
Test Scenario
OSL Result
Expected
Confidence
Sally-Anne Basic
PASS
PASS
1.000
Sally-Anne with Distractor
PASS
PASS
1.000
Nested Belief (Level 2)
PASS
PASS
0.950
Multiple Objects
PASS
PASS
1.000
Temporal Belief Change
PASS
PASS
0.975
False Photograph Task
PASS
PASS
1.000
Appendix
Table 5. Theory-of-mind test results across multiple scenarios. PASS indicates that the produced result matches the expected outcome for the corresponding scenario.
Figure 4. Illustrative examples of observer–situation lattice structures. Nodes are labeled by (o,σ) pairs; edges depict order relations.
Lattice
Observers
Situations
Balance
Balanced
Avg Time
Std Dev
Affected
Max
Trials
Size n
∣O∣
∣Σ∣
Factor κ
( μ s)
( μ s)
Elements
Affected
9
3
3
1.00
Yes
8.31
7.07
2.5 ± 2.3
8
30
16
4
4
1.00
Yes
9.79
7.84
3.5 ± 3.1
11
30
30
5
6
1.20
Yes
21.21
19.01
7.1 ± 6.9
23
30
56
7
8
1.14
Yes
13.91
13.57
4.8 ± 5.4
17
30
100
10
10
1.00
Yes
9.25
9.53
3.2 ± 3.7
19
30
Appendix
Table 6. Extended RBP scaling performance results with balance factors. All values are from measured runs of our implementation; the number of trials per configuration is reported in the Trials column.
Method
Avg Time ( μ s)
Memory (MB)
Empirical scaling (fit)
Max Size
OSL (Ours)
55.94 ± 94.57
12.3
n0.336
100,000+
ATMS
847.2 ± 234.1
28.7
n1.2
5,000
Distributed TMS
167.3 ± 89.4
49.6
n0.8
10,000
MEPK Planner
2341.7 ± 1456.2
156.4
(rapid growth)
50 agents
Appendix
Table 7. Baseline comparison results on lattices up to n≤2000 elements. “Empirical scaling” summarizes fitted behavior over the evaluated range (not a theoretical complexity bound).
Large Language Model (LLM) agents deployed in complex real-world scenarios increasingly operate as spatially distributed entities. However, this physical dispersion constrains agents to limited local perception and finite temporal horizons. We characterize this bottleneck as spatiotemporal partial observability, where spatial and temporal limitations are fundamentally coupled: resolving spatial conflicts requires temporal reasoning about neighbors' future actions, while temporal planning requires spatial context beyond local perception. To bridge this gap, we introduce MACRO-LLM, LLM-empowered multi-agent collaborative reasoning under spatiotemporal partial observability. The architecture interleaves spatial and temporal reasoning within each decision cycle via three interdependent modules: (1) the CoProposer mitigates temporal uncertainty by verifying candidate actions via predictive rollouts; (2) the Negotiator overcomes spatial myopia by resolving conflicts through mean-field statistical aggregation, grounded in the CoProposer's rollout rewards; and (3) the Introspector closes the reasoning loop by analyzing environmental drift and attributing performance changes to refine strategies. Extensive evaluations on two complex long-horizon tasks, cooperative platoon planning and pandemic control, demonstrate that our framework enables robust coordination under spatiotemporal partial observability.
Large Language Model agents have rapidly evolved from static text generators into dynamic systems capable of executing complex autonomous workflows. To enhance reliability, multi-agent frameworks assigning specialized roles are increasingly adopted to enable self-reflection and mutual auditing. While such role-playing effectively leverages domain expert knowledge, we find it simultaneously induces a human-like cognitive bias known as Actor-Observer Asymmetry (AOA). Specifically, an agent acting as an actor (during self-reflection) tends to attribute failures to external factors, whereas an observer (during mutual auditing) attributes the same errors to internal faults. We quantify this using our new Ambiguous Failure Benchmark, which reveals that simply swapping perspectives triggers the AOA effect in over 20% of cases for most models. To tame this bias, we introduce ReTAS (Reasoning via Thesis-Antithesis-Synthesis), a model trained through dialectical alignment to enforce perspective-invariant reasoning. By integrating dialectical chain-of-thought with Group Relative Policy Optimization, ReTAS guides agents to synthesize conflicting viewpoints into an objective consensus. Experiments demonstrate that ReTAS effectively mitigates attribution inconsistency and significantly improves fault resolution rates in ambiguous scenarios.
Bobo Li, Rui Wu, Zibo Ji +5
National University of Singapore · Sichuan University · University of Minnesota Twin Cities +2
Recently Large Language Models (LLMs) have been increasingly deployed as autonomous agents in applications such as self-reflection, retrieval-augmented generation, and scientific discovery. In these settings, agents must act based on limited observations rather than full environmental states, leading to partial observability. This introduces several key challenges: belief state inference, task objective misalignment, and planning under uncertainty. Prior approaches typically condition actions on full or summarized action-observation histories whose redundant and irrelevant information can mislead the decision making of LLM agent. Inspired by human cognition, we propose a novel neuro-symbolic fast-slow thinking (NeSyFS) framework for LLM agent, addressing the challenges introduced by partial observability in a unified approach. We use a knowledge graph (KG) to represent the belief state, providing triplets as context for every module of NeSyFS. The fast-thinking module performs reactive action, while slow-thinking conducts a new uncertainty-aware planning by following the high-level structure of twisted sequential Monte Carlo (TSMC) algorithm. To mitigate the misalignment of task objective, a reflection module is used to reflect fast-thinking actions, and also switches to the slow-thinking module whenever reactive actions repeatedly fail. Experiments on three representative benchmarks, i.e. ALFWorld, Webshop, and ScienceWorld, demonstrate significant advantages over previous methods.
Duo Xu, Faramarz Fekri
Georgia Institute of Technology 225 North Ave NW, Atlanta, GA 30332