acceptodds
Under review as a conference paper at ICLR 2027

Language Models Track State but Underuse Observation Reliability

Abstract

Reliable state estimation requires adjusting each update to the quality of new evidence. We study this adjustment in tasks with exact Bayesian references. A five-model survey reveals strong posterior-mean tracking but weak adaptation to sensor reliability. We introduce a paired audit that holds the displayed history fixed and perturbs only the current reading, measuring its influence without regressing on previous model outputs. Under greedy decoding, three checkpoints from two families realize only 5-8% of the prescribed gain change in a matched perturbation test as sensor variance increases sixteenfold. Focused experiments reveal that the same checkpoint can produce accurate updates: supplying the predicted state improves posterior accuracy, and a reasoning-enabled protocol sharply reduces both local-gain and posterior-mean errors on held-out trajectories with unchanged prompts. We translate this diagnosis into the Perception-Dynamics Decoupled Agent (PDDA), combining language-based report extraction with explicit Kalman filtering. Complete-system and matched-extraction comparisons improve monitoring accuracy without generated reasoning at the numerical update stage. Reliability-conditioned updating thus provides a diagnostic beyond trajectory agreement and a concrete target for improving language-based estimation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
large language models, Bayesian filtering, Kalman filtering, state estimation, observation reliability, uncertainty quantification, belief updating

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xRpVOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper distinguishes between "posterior mean tracking" and "adaptive adjustment for observation reliability." The main diagnostic method is to apply a symmetrical perturbation to the current reading while keeping the displayed historical data unchanged, and then compare the finite difference response with the exact Kalman gain under different observation variance settings. Through three checkpoints, approximately 5%–8% of the required endpoint gain change under the baseline decoding condition can be achieved. Through three specific checkpoints, it is possible to achieve approximately 5%–8% of the required adjustment in terminal gain under baseline decoding conditions. A "held-out reasoning protocol" is shown to improve both local responses and the posterior mean. Furthermore, this paper evaluates a system that extracts linguistic reports and applies an explicit Kalman filter, providing a comparative analysis based on the same stream of extracted data.

Critical strengths
  1. A well-defined behavioral diagnostic analysis (see Equations (1)–(2), Sections 2–3, and Appendix B.1) avoids the "shared previous-report error" that can arise in aggregate "effective gain" regression analyses by fixing historical values ​​and employing context-specific exact gains.

  2. The study conducted "recovery" tests using the same output results and held-out trajectories.

Critical weaknesses
  1. The contribution is substantially narrower than the title and system framing suggest. The strongest paired result covers three relatively small checkpoints from two families on scalar linear-Gaussian filtering; successful reasoning recovery is demonstrated on one checkpoint. Existing work already studies Bayesian belief trajectories and discounted filtering. PDDA itself is a language extraction interface followed by a classical estimator, and it exactly matches the generated-code Kalman baseline. or a general limitation of language models. Evidence across more model families and nonlinear settings would materially strengthen significance.
  2. Some secondary evidence does not identify the mechanism or unconditional capability being discussed. The budget sweep omits complete-pair denominators and treatment of failed pairs. A successful long-budget reasoning protocol therefore does not establish performance at a fixed computational budget. Moreover, schedule GAR divides innovation-energy-weighted fitted gains by arithmetic phase-mean reference gains ; an exact time-varying filter need not self-score to one. These are acknowledged limitations, but the secondary plots must not be read as calibrated mechanistic evidence. Steering readout shifts and oracle-state assistance likewise cannot establish where the original computation fails.
Questions for authors
  1. Could you report the full pairwise validity and unconditional error for each budget, as well as a factor-based comparison—distinguishing between "thinking" modes, temperature, and token budgets—within the held-out scenarios?

  2. Can the reliability-response gap be replicated using the same pairwise estimator and comparable prompts across other independently developed model families or in non-linear filtering tasks?

  3. Regarding the comparison of switching noise, what are the GAR (Generalized Accuracy/Agreement Rate) values ​​obtained using an exact filter, given the same regression model and denominator? Would using a denominator based on "innovation energy matching" alter the conclusions?

  4. Can the existing data reproduce all the primary pairwise statistics found in the original invocation dataset? Additionally, could you correct the author list for the reference to Liu et al. provided below?

6 comments
AuthorsOpenReview ↗

Response to Reviewer xRpV (1/4)

We sincerely thank the reviewer for the careful and detailed comments, and we appreciate your recognition of the fixed-history design and of the recovery tests on held-out trajectories. For easier reference, we label the points in W1 as W1-1 to W1-4 and those in W2 as W2-1 to W2-3, and answer Q1 to Q4 together with the related points. We have uploaded a revised PDF in which all changes are marked in blue.

W1-1 & Q2. More model families and a nonlinear task

The models barely adjust their weight on a new reading to its reliability, and this gap is replicated on other model families and on a nonlinear filtering task.

Two quantities are used below. The weight (Kalman gain) is the share that a model gives to the new reading when it updates its estimate, and the optimal weight decreases as the sensor noise variance R increases. GAR is the decrease in the model's weight as R goes from 1 to 16, divided by the decrease in the optimal weight. It is 1 for a model that adjusts optimally and 0 for a model that does not adjust at all.

Following your suggestion in Q2, we evaluated five additional models with the same paired estimator, prompts, and settings as in Table 1B (δ = 3, 120 paired contexts per declared variance, greedy non-thinking decoding). Together with the three original checkpoints, the audit now covers eight models from four families (Qwen, Mistral, Llama, and OLMo), ranging from 1.7B to 32B parameters.

Table R1. Weights and GAR of eight models at three noise variances (Table 28, App. N.1 of the revised PDF).

ModelR = 1R = 4R = 16GAR95% CI
Exact Kalman filter0.6180.3910.2211—
Qwen3-8B0.6400.6380.6190.054[0.008, 0.101]
Qwen3-4B0.5840.5720.5610.058[0.010, 0.108]
Mistral-7B0.6820.6700.6510.078[0.020, 0.139]
Qwen3-1.7B0.1920.1790.1830.024[−0.038, 0.108]
OLMo-2-7B-Instruct0.3750.3680.3560.050[0.014, 0.089]
Llama-3.1-8B-Instruct0.2270.2130.2150.029[−0.040, 0.111]
Qwen3-30B-A3B0.5390.5210.5260.033[−0.037, 0.114]
Qwen3-32B0.4000.3950.3850.040[0.016, 0.065]

As R increases from 1 to 16, the optimal weight decreases from 0.618 to 0.221, whereas the weights of all eight models remain almost unchanged. All GARs are below 0.08, and even the largest upper end of the 95% CIs is only 0.139, so the gap is replicated in all four model families. Larger models do not reduce the gap either, since the GARs of Qwen3-32B (0.040) and Qwen3-30B-A3B (0.033) are no higher than that of Qwen3-8B (0.054). To make sure that the new runs are directly comparable with the original ones, we also re-ran Qwen3-8B and obtained 0.057 [0.012, 0.105], in agreement with the 0.054 of Table 1B.

For the nonlinear task, we replaced the linear dynamics with xt=xt−1+2sin⁡(xt−1/2)+wtx_t = x_{t-1} + 2\sin(x_{t-1}/2) + w_t and kept all other settings unchanged. The optimal weights are computed by a grid filter with 2,001 points. Table R2 reports four models from three families on this task.

Table R2. Weights and GAR of four models in the nonlinear task (Table 29, App. N.2 of the revised PDF).

ModelR = 1R = 4R = 16GAR95% CI
Exact grid filter0.5150.2140.0721—
Qwen3-8B0.8200.8370.834−0.030[−0.077, 0.006]
Qwen3-32B0.5230.5170.5060.038[0.009, 0.069]
Llama-3.1-8B-Instruct0.2980.2870.2850.029[−0.041, 0.102]
OLMo-2-7B-Instruct0.4710.4620.4530.041[0.004, 0.081]

In this task the optimal weight decreases from 0.515 to 0.072, while the weight of each of the four models changes by less than 0.02 and the largest GAR is only 0.041. The gap is therefore not specific to linear-Gaussian filtering, which is also in line with the discrete HMM result in Table 20 of the paper.

Taken together, both statements in the title now rest on broader evidence. "Track state" means that the estimates follow the state, and the estimates of the five surveyed models correlate with the exact posterior mean at 0.927 to 0.977 (Table 1A). "Underuse observation reliability" means that the weight is not adjusted to the reliability, which now holds for eight models from four families and for the nonlinear task. We have added these results to Section 3 and to the new Apps. N.1 and N.2 of the revised PDF.

AuthorsOpenReview ↗

Response to Reviewer xRpV (2/4)

W1-2. Reasoning recovery shown on one checkpoint

Reasoning sharply reduces the weight error, and this now holds for five models from three families. We repeated the held-out experiment of Table 5 on four more models and compared the weight error (local-gain MAE, the mean absolute difference between the model's weight and the optimal weight) under the default and the reasoning settings. Qwen3-32B and Qwen3-30B-A3B enter the reasoning setting by enabling thinking, as Qwen3-8B does. Llama-3.1-8B-Instruct and OLMo-2-7B-Instruct have no thinking mode, so one sentence appended to the otherwise unchanged prompt asks them to write out the calculation step by step before the answer, with a token cap large enough for the full calculation.

Table R3. Weight error of five models under the default and the reasoning settings (Table 30, App. N.3 of the revised PDF).

ModelDefaultReasoning
Qwen3-8B0.3010.019
Qwen3-32B0.1790.017
Qwen3-30B-A3B0.2250.082
Llama-3.1-8B-Instruct0.2470.038
OLMo-2-7B-Instruct0.1860.044

Under the reasoning setting, the weight error drops by more than 60% in all five models and falls below 0.05 in four of them, so the recovery under reasoning is not limited to Qwen3-8B. Qwen3-30B-A3B, a mixture-of-experts model that activates about 3B parameters per token, improves from 0.225 to 0.082.

W1-3. Difference from prior work on belief trajectories and discounted filtering

Prior work compares how close a model's whole trajectory of estimates is to the Bayes-optimal trajectory, whereas our audit measures whether the weight that a model gives to a new reading changes with the reliability of the sensor. The latter cannot be seen from trajectory agreement. The reason is that a model that gives every new reading roughly the same weight can still stay close to the optimal estimates on a trajectory with a constant noise level (Section 1). This is what we observe. The estimates of all five models correlate with the exact posterior mean at 0.927 or above (Table 1A), yet when the declared noise variance increases from 1 to 16, the optimal weight should decrease from 0.618 to 0.221, and the weights of the three checkpoints in Table 1B decrease by only 5% to 8% of this amount. In other words, the trajectory agrees well while the weight barely adapts to reliability. The fixed-history audit is designed to measure this weight directly. It shows the model exactly the same history of readings, changes only the noise variance declared in the prompt, and shifts the current reading up and down by δ to see how much the estimate follows. This gives the weight of that single update, which is compared with the exact Kalman gain under the same conditions. As you noted among the strengths, this design also avoids the shared previous-report error of aggregate regressions.

Studies of discounted filtering fit a model's trajectory of estimates with a fixed discount factor. A fixed discount does correspond to a weight that does not change with the noise, but it cannot explain the uncertainty that the model reports. As the noise variance increases from 1 to 16, a fixed discount predicts that the interval widens by a factor of 4 and the exact filter by 2.39, while the model's interval widens by only 1.79 (Section 3, App. B.6). A fixed discount therefore describes the trajectory of estimates but not the complete update of the model. We discuss these studies in Section 7.

W1-4. PDDA and the generated-code Kalman baseline

PDDA and the generated-code Kalman baseline have the same accuracy, which is expected because both carry out the numerical update with the Kalman formula. The difference between them is that PDDA needs no generation for an update. The four updaters compared in Table 2B receive exactly the same readings, which the language model has extracted from the reports, and they differ only in who carries out the update from these readings. When the language model performs the update itself, the estimation error is 0.435, and it is 0.038 when the model reasons first. Whenever the update uses the Kalman formula, either executed by model-generated code or by PDDA, the error is 0.028. This shows that the error arises when the language model performs the numerical update itself, which agrees with the conclusion of the fixed-history audit. Between the two options that use the Kalman formula, the code tool still has to generate about 180 tokens of code per update and takes about 0.45 s, whereas PDDA generates nothing and takes less than 1 ms (Table 6). The role of PDDA in the paper is therefore to show that this finding translates directly into a system design, in which a filter performs the numerical update and the language model only reads the values from the reports.

AuthorsOpenReview ↗

Response to Reviewer xRpV (3/4)

W2-1 & Q1-1. Budget sweep: complete pairs and the unconditional error

Following your comment, we have replaced the original budget sweep with one that reports every denominator. The results show that reasoning helps only when the token budget is large enough for it to finish.

As you pointed out, the budget sweep in App. B.5 was run on the development set, computed errors over completed pairs only, and did not record how many pairs completed, so its values cannot represent unconditional performance. We therefore re-ran the sweep on the 12 held-out trajectories of Table 5 (Qwen3-8B, thinking enabled, temperature 0.6) and used it to replace the original values in Section 5 and App. B.5. Each weight is computed from a pair of calls, in which the current reading is shifted up and down by δ, and a pair is complete only when both calls return a parsable answer. Each budget has 864 calls and 432 pairs. Pairs that fail to complete are not discarded. They are scored as if the model had not updated its estimate, that is, with a weight of zero. The last column of Table R4 is therefore the unconditional error requested in Q1.

Table R4. Budget sweep on the held-out trajectories (Table 32, App. N.5 of the revised PDF).

Token capParsed callsComplete pairsWeight error, complete pairsWeight error, all pairs
2,0480%0 of 432—0.410
4,0961%0 of 432—0.410
8,19256%147 of 4320.0140.282
16,384100%432 of 4320.0110.011
32,768100%432 of 4320.0190.019

The reasoning trace of the model has 8,066 tokens on average (App. B.3), so no pair completes with a cap of 4,096 tokens or fewer, and the unconditional error is 0.410. With 16,384 tokens or more, all pairs complete and the error is between 0.011 and 0.019, so reasoning helps only when it can finish. The last row is the protocol of Table 5.

Success at a long budget does not establish performance at a small fixed budget, and Table R4 shows that reasoning does not help when the budget is small. This is the motivation for PDDA, whose update does not generate any tokens.

One additional note on scoring. Truncated calls are counted as unparsed in Table R4. If they are instead parsed with the parser of App. M, only the 8,192 row changes, from 147 to 151 complete pairs.

Q1-2. Separating thinking mode, temperature, and token budget

Among the three factors, temperature has little effect, and the weight error drops only when thinking is enabled with a sufficient token budget. We changed one factor at a time on the held-out trajectories of Table 5, and the results are given in Table R5. The scoring is the same as in Table R4, so pairs that fail to complete are included in the error.

Table R5. One factor at a time on the held-out trajectories (Table 31, App. N.4 of the revised PDF).

ThinkingTemperatureToken capComplete pairsWeight error, all pairs
off0512432 of 4320.301
off0.6512432 of 4320.287
on032,768427 of 4320.024
on0.62,0480 of 4320.410
on0.632,768432 of 4320.019

Without thinking, increasing the temperature from 0 to 0.6 changes the error only from 0.301 to 0.287, so temperature has little effect. In this case all calls finish within 512 tokens, so the cap is not binding and larger caps were not tested. With thinking and 32,768 tokens, the error is about 0.02 at both temperatures. With thinking but only 2,048 tokens, the reasoning cannot finish and the error is 0.410, so the improvement requires thinking with a sufficient budget, and our statements about reasoning in the paper refer to this complete setting. The first and last rows are the two protocols of Table 5.

Reviewer_LSpzOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2(−1)
Soundness
2
Read review
Paper summary

This paper investigates whether large language models appropriately adapt their state estimates to observation reliability. The authors introduce a fixed-history reliability audit to measure how model predictions respond to changes in sensor uncertainty. Experiments show that LLMs can accurately track posterior means while substantially underusing reliability information. The paper further demonstrates that reasoning-enabled inference improves estimation accuracy and proposes the Perception-Dynamics Decoupled Agent (PDDA), which combines LLM-based observation extraction with explicit Kalman filtering.

Critical strengths
  1. The paper highlights an important distinction between accurate state tracking and reliability-aware belief updating, providing useful insights into LLM behavior.
  2. The fixed-history perturbation approach directly measures observation sensitivity while reducing confounding effects from previous model predictions.
  3. The experiments include multiple language models, controlled reliability interventions, reasoning-protocol comparisons, and statistical confidence intervals.
Critical weaknesses
  1. Most experiments rely on one-dimensional linear Gaussian systems with known dynamics and observation noise. These settings do not adequately demonstrate generalization to nonlinear dynamics, unknown uncertainty, or realistic agent environments.
  2. The reasoning-enabled protocol dramatically improves estimation accuracy, suggesting that the observed deficiencies may result from numerical computation limitations rather than an inability to understand observation reliability. Since reasoning mode, temperature, and computational budget change simultaneously, the experiments cannot isolate the underlying cause.
  3. The main fixed-history conclusions are established primarily on relatively small models under non-thinking inference. Evaluation on larger reasoning models and more diverse inference settings is necessary to establish whether the identified limitations remain relevant to modern LLM-based agents.
Questions for authors

NA

6 comments
AuthorsOpenReview ↗

Response to Reviewer LSpz (1/3)

We sincerely thank the reviewer for the constructive comments, and we appreciate your recognition of the distinction between accurate state tracking and reliability-aware updating, as well as of the fixed-history design. We have uploaded a revised PDF in which all changes are marked in blue.

W1. Beyond one-dimensional linear-Gaussian systems

Our findings also hold beyond one-dimensional linear-Gaussian systems. Of the three aspects you mention, we have added an experiment on nonlinear dynamics, and the paper already contains results on unknown uncertainty and on more realistic environments. We use the linear-Gaussian system in the main experiments because every update has a closed-form optimal weight in this setting, so the deviation of a model can be quantified exactly. Here the weight (Kalman gain) is the share that a model gives to the new reading when it updates its estimate, and the optimal weight decreases as the sensor noise variance R increases.

(1) Nonlinear dynamics. We replaced the linear dynamics with xt=xt−1+2sin⁡(xt−1/2)+wtx_t = x_{t-1} + 2\sin(x_{t-1}/2) + w_t, kept the other settings of Table 1B, and computed the optimal weights with a grid filter. Table R1 reports four models from three families. GAR is the decrease in the model's weight as R goes from 1 to 16, divided by the decrease in the optimal weight. It is 1 for a model that adjusts optimally and 0 for a model that does not adjust at all.

Table R1. Weights and GAR of four models in the nonlinear task (Table 29, App. N.2 of the revised PDF).

ModelR = 1R = 4R = 16GAR95% CI
Exact grid filter0.5150.2140.0721—
Qwen3-8B0.8200.8370.834−0.030[−0.077, 0.006]
Qwen3-32B0.5230.5170.5060.038[0.009, 0.069]
Llama-3.1-8B-Instruct0.2980.2870.2850.029[−0.041, 0.102]
OLMo-2-7B-Instruct0.4710.4620.4530.041[0.004, 0.081]

The optimal weight decreases from 0.515 to 0.072, while the weight of each of the four models changes by less than 0.02 and the largest GAR is only 0.041, so the models also fail to adjust their weight to reliability under nonlinear dynamics. In addition, in the discrete five-room HMM (Table 20), the belief of the model moves only 0.435 times as much as the exact posterior under the more accurate sensor.

(2) Unknown uncertainty. Section 4 tests the case where the actual noise changes without being announced to the model. The GAR is −0.098 in this case, which means that the model does not adjust its weight from the data alone. PDDA, the system proposed in the paper, lets the language model only read the values from the reports and assigns the numerical update to a Kalman filter. When PDDA assumes a sensor variance that is one third of or 3 times the true value, its estimation errors are 0.265 and 0.114, still below the 0.471 obtained when the language model produces the estimate directly (Table 26).

(3) More realistic settings. In the telemetry monitoring task, where the inputs are natural-language reports, the error is 0.471 when the language model produces the estimate directly and 0.017 for PDDA (Table 2A). On the real Mauna Loa CO₂ series, PDDA reaches an error of 0.605, close to the 0.609 of a Kalman filter that uses the original readings, while the language model alone has 0.723, so explicit filtering remains effective on real data.

The failure to adjust the weight to reliability is therefore not specific to linear-Gaussian systems, and assigning the numerical update to a filter remains effective on real data and when the noise level is not announced accurately. We have added the nonlinear task to Section 3 and to the new App. N.2 of the revised PDF.

AuthorsOpenReview ↗

Response to Reviewer LSpz (2/3)

W2. Numerical computation or understanding of reliability, and separating the three factors

The problem lies in the numerical computation and not in the model's understanding of observation reliability. The title says "underuse" rather than "cannot use" to express the same point. Three results support this.

  1. The model carries the reliability information internally. A residual-stream probe decodes the noise level declared in the prompt (R2=0.978R^2 = 0.978 within the seen levels, App. I).
  2. Providing the formulas alone does not solve the problem. Without thinking, the weight of the model stays between 0.47 and 0.54 even when formulas and worked examples are given in the prompt, far above the optimal value of 0.221 at R=16R = 16 (Table 16).
  3. The update is accurate once the intermediate result is given to the model. In the staged experiment with reasoning enabled, additionally providing the predicted mean and variance reduces the posterior-mean error from 0.0226 to 0.000077 (Table 4).

Regarding your concern that the three factors change together, we ran an additional experiment on the held-out trajectories of Table 5, changing one factor at a time. The error in the table is the weight error (local-gain MAE), the mean absolute difference between the model's weight and the optimal weight. Each weight is computed from a pair of calls, a pair is complete only when both calls return a parsable answer, and pairs that fail to complete enter the error with a weight of zero. Each condition has 432 pairs.

Table R2. One factor at a time on the held-out trajectories (Table 31, App. N.4 of the revised PDF).

ThinkingTemperatureToken capComplete pairsWeight error
off0512432 of 4320.301
off0.6512432 of 4320.287
on032,768427 of 4320.024
on0.62,0480 of 4320.410
on0.632,768432 of 4320.019

As shown in Table R2, temperature has little effect, and the error drops clearly only when thinking is enabled and the token budget is large enough for the reasoning to finish. Without thinking, all calls finish within 512 tokens, so the cap is not binding and larger caps were not tested. In other words, the model computes the update correctly only when it has enough room to write out the calculation step by step, so the bottleneck is the numerical computation itself. This is also the reason why PDDA assigns the numerical update to a filter and uses the language model only to read the reports. The first and last rows of Table R2 are the two protocols of Table 5.

AuthorsOpenReview ↗

Response to Reviewer LSpz (3/3)

W3. Larger models and reasoning-enabled inference

The failure to adjust the weight to reliability remains in larger models. Under the reasoning setting, the weight error drops sharply in all five models from three families, but at a high cost per update.

For larger models, we evaluated two larger models and one new model family with the settings of Table 1B.

Table R3. GAR of the larger models and the new family (four rows of Table 28, App. N.1 of the revised PDF).

ModelGAR95% CI
Qwen3-8B0.054[0.008, 0.101]
Llama-3.1-8B-Instruct0.029[−0.040, 0.111]
Qwen3-30B-A3B0.033[−0.037, 0.114]
Qwen3-32B0.040[0.016, 0.065]

None of the new models has a higher GAR than the three checkpoints of Table 1B (0.054 to 0.078), so the gap remains as the models become larger. Table 14 of the paper shows the same behavior for Qwen3-30B-A3B. While the declared variance changes 128-fold and the optimal weight decreases from 0.83 to 0.16, the weight of the model stays between 0.41 and 0.44.

For reasoning-enabled inference, we compared the weight error under the default and the reasoning settings on the 12 held-out trajectories of Table 5. The default setting uses no thinking, temperature 0, and at most 512 tokens. For the three Qwen3 models, the reasoning setting enables thinking with temperature 0.6 and at most 32,768 tokens. Llama-3.1-8B-Instruct and OLMo-2-7B-Instruct have no thinking mode, so one sentence appended to the otherwise unchanged prompt asks them to write out the calculation step by step before the answer, with a token cap large enough for the full calculation.

Table R4. Weight error of five models under the default and the reasoning settings (Table 30, App. N.3 of the revised PDF).

ModelDefaultReasoning
Qwen3-8B0.3010.019
Qwen3-32B0.1790.017
Qwen3-30B-A3B0.2250.082
Llama-3.1-8B-Instruct0.2470.038
OLMo-2-7B-Instruct0.1860.044

Under the reasoning setting, the weight error drops by more than 60% in all five models and falls below 0.05 in four of them, so larger models and models from other families also recover through reasoning. Qwen3-30B-A3B, a mixture-of-experts model that activates about 3B parameters per token, improves from 0.225 to 0.082. However, reasoning is expensive. Qwen3-32B generates about 6,600 tokens per call on average, and Table 6 of the paper reports about 7,850 tokens and 14.2 s for one reasoning update of Qwen3-8B. This is a considerable cost for an agent that updates its state at every step, whereas the update of PDDA does not generate any tokens. The finding therefore remains relevant to current LLM-based agents. The results under other temperatures and token caps are given in Table R2 under W2. We have added these results to Sections 3 and 5 and to the new App. N of the revised PDF.

Reviewer_geJKOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates whether language models appropriately adjust their state estimates according to observation reliability. Using tasks with exact Bayesian references, it shows that good agreement with state trajectories can coexist with insufficient adaptation to sensor reliability under the base inference protocol. The paper introduces a fixed-history evaluation to isolate this behavior, examines recovery through supplied state information and reasoning, and evaluates a system combining language-based observation extraction with explicit Kalman filtering.

Critical strengths
  1. A clear question supported by well-controlled evaluation. Holding histories fixed helps isolate how declared observation reliability affects the model’s updates. This directly addresses the ambiguity that good trajectory tracking may conceal inappropriate evidence weighting. The consistency across multiple checkpoints strengthens the central finding.

  2. The diagnosis is connected to concrete improvements. The reasoning experiments demonstrate improvements in both update responses and estimated means. The comparisons using identical extracted observations further separate report interpretation from numerical estimation. Together, these results provide useful evidence about how to improve performance and the associated computational costs.

Critical weaknesses
  1. Limited algorithmic novelty. The main contribution lies in the targeted evaluation and empirical findings. PDDA combines language-based extraction with an established Kalman filter, so its algorithmic novelty is limited. The paper’s significance therefore depends primarily on how much the proposed diagnosis advances our understanding beyond earlier studies of evidence reliability and belief updating.

  2. The generality of the findings remains limited. The central paired audit covers three checkpoints from two model families, while the detailed reasoning recovery focuses on Qwen3-8B. These results support the conclusions under the tested settings, but leave uncertainty about how widely the same limitation occurs in larger models and reasoning-enabled use. This limits the breadth of the practical implications.

Questions for authors
  1. Compared with prior studies of evidence reliability, what is the main new insight provided by the fixed-history audit?

  2. Does the observed limitation also occur in larger models and reasoning-enabled settings beyond Qwen3-8B?

5 comments
AuthorsOpenReview ↗

Response to Reviewer geJK (1/2)

We sincerely thank the reviewer for the positive and constructive comments, and we appreciate your recognition of the clear question, the well-controlled evaluation, and the connection between the diagnosis and concrete improvements. We have uploaded a revised PDF in which all changes are marked in blue.

W1 & Q1. The new insight of the fixed-history audit, and the role of PDDA

The new insight of the fixed-history audit is that a model can track the state accurately while the weight it gives to a new reading barely changes with the reliability of that reading. Here the weight (Kalman gain) is the share that a model gives to the new reading when it updates its estimate, and the optimal weight decreases as the sensor noise increases. Prior studies mainly compare how close a model's trajectory of estimates is to the Bayes-optimal trajectory, which cannot reveal this, because a model that gives every new reading roughly the same weight can still track well on a trajectory with a constant noise level (Section 1). The fixed-history audit shows the model exactly the same history of readings and changes only the noise variance declared in the prompt. It therefore measures the weight of a single update directly and compares it with the exact Kalman gain, without regressing on the model's own earlier outputs. As a result, the estimates of all five models correlate with the exact posterior mean at 0.927 or above (Table 1A), yet when the noise variance increases from 1 to 16, the optimal weight should decrease from 0.618 to 0.221, and the weights of the three checkpoints in Table 1B decrease by only 5% to 8% of this amount. Accurate tracking therefore does not imply that the weight adapts to reliability. The revised Introduction now states this insight explicitly in one sentence.

Regarding PDDA, its role in the paper is to show that these findings translate directly into a system design. Since the problem lies in the numerical update, PDDA assigns this step to a Kalman filter and keeps the language model for reading the reports, where it is strong (recall 0.989, against 0.866 for the audited regex, Section 6.3). With this design, the estimation error of the complete system decreases from 0.471 to 0.017 (Table 2A).

AuthorsOpenReview ↗

Response to Reviewer geJK (2/2)

W2 & Q2. Larger models and reasoning-enabled use

The failure to adjust the weight to reliability also occurs in larger models. Under the reasoning setting, the weight error of the other models also drops sharply, as it does for Qwen3-8B.

We extended the audit of Table 1B from three models to eight, covering four families and sizes from 1.7B to 32B. The GAR in Table R1 is the decrease in the model's weight as the noise variance goes from 1 to 16, divided by the decrease in the optimal weight. It is 1 for a model that adjusts optimally and 0 for a model that does not adjust at all.

Table R1. GAR of eight models (Table 28, App. N.1 of the revised PDF).

ModelGAR95% CI
Qwen3-8B0.054[0.008, 0.101]
Qwen3-4B0.058[0.010, 0.108]
Mistral-7B0.078[0.020, 0.139]
Qwen3-1.7B0.024[−0.038, 0.108]
OLMo-2-7B-Instruct0.050[0.014, 0.089]
Llama-3.1-8B-Instruct0.029[−0.040, 0.111]
Qwen3-30B-A3B0.033[−0.037, 0.114]
Qwen3-32B0.040[0.016, 0.065]

All eight models have a GAR below 0.08, and the five new ones are at or below 0.05. They include the two larger models, Qwen3-30B-A3B and Qwen3-32B, so the limitation also occurs in larger models.

For reasoning-enabled use, we repeated the held-out experiment of Table 5 on four more models and compared the weight error (local-gain MAE, the mean absolute difference between the model's weight and the optimal weight) under the default and the reasoning settings. Qwen3-32B and Qwen3-30B-A3B enter the reasoning setting by enabling thinking, as Qwen3-8B does. Llama-3.1-8B-Instruct and OLMo-2-7B-Instruct have no thinking mode, so one sentence appended to the otherwise unchanged prompt asks them to write out the calculation step by step before the answer, with a token cap large enough for the full calculation.

Table R2. Weight error of five models under the default and the reasoning settings (Table 30, App. N.3 of the revised PDF).

ModelDefaultReasoning
Qwen3-8B0.3010.019
Qwen3-32B0.1790.017
Qwen3-30B-A3B0.2250.082
Llama-3.1-8B-Instruct0.2470.038
OLMo-2-7B-Instruct0.1860.044

Under the reasoning setting, the weight error drops by more than 60% in all five models and falls below 0.05 in four of them. Qwen3-30B-A3B, a mixture-of-experts model that activates about 3B parameters per token, improves from 0.225 to 0.082. This is what the title means by "underuse": the models are able to give a reading the correct weight, but they do not do so under the default setting. Reasoning also generates thousands of tokens per update (about 7,850 in Table 6 of the paper), a cost that PDDA does not incur. We have added these results to Sections 3 and 5 and to the new App. N of the revised PDF.

Dear Reviewer geJK,

Thanks for your time and effort in reviewing this paper. As explained in the rebuttal, the new insight is that a model can estimate the state accurately while weighting accurate and noisy readings almost equally (W1, Q1). This also occurs in larger models of up to 32B, and enabling reasoning largely corrects it in all five tested models (W2, Q2).

As the discussion period will end in about 3 days, we gently request that you check whether these address your concerns. Thanks again for your time and help in improving the quality of this paper.

Regards from the authors.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their time and effort in providing careful and constructive feedback on our submission.


We are encouraged to see that our reviewers recognize the fixed-history audit as a well-defined and well-controlled diagnostic.

  • Reviewer geJK notes a "clear question supported by well-controlled evaluation", and Reviewer xRpV describes the audit as a "well-defined behavioral diagnostic analysis".
  • Reviewer LSpz notes that the paper "highlights an important distinction between accurate state tracking and reliability-aware belief updating" and that the fixed-history approach "directly measures observation sensitivity".
  • Reviewer geJK notes that the diagnosis "is connected to concrete improvements", and Reviewer LSpz notes that the experiments "include multiple language models, controlled reliability interventions, reasoning-protocol comparisons, and statistical confidence intervals".

We highlight our contributions as follows:

  • Our Fixed-History Paired Audit holds the displayed history fixed, perturbs only the current reading, and compares the weight that a model gives to this reading with the Exact Kalman Gain. It identifies a gap that trajectory agreement alone can conceal: a model can track the posterior mean accurately while, under the base protocol, this weight barely changes with the reliability of the reading.
  • We further show that the same checkpoint can produce accurate updates under a Reasoning-Enabled Protocol with a sufficient token budget, and translate the diagnosis into the Perception-Dynamics Decoupled Agent (PDDA), which combines Language-Based Report Extraction with Explicit Kalman Filtering and generates no tokens for the numerical update.

As suggested by the reviewers, we revised our manuscript and added Appendix Section N, where the tables in our responses appear as Table 28 through Table 33. Each table in our responses is labeled with its number in the revised PDF, and new and revised text is marked in blue.

Additional Results & Experiments

  • Added the fixed-history audit on five more models, so that it covers eight models from four families and sizes from 1.7B to 32B, in Table 28 and Appendix Section N.1, as suggested by Reviewers xRpV, LSpz, and geJK.
  • Added a nonlinear filtering task, with reference weights from a grid filter, in Table 29 and Appendix Section N.2, as suggested by Reviewers xRpV and LSpz.
  • Added the comparison of the base and reasoning protocols on additional models, evaluated on the held-out trajectories, in Table 30 and Appendix Section N.3, as suggested by Reviewers xRpV, LSpz, and geJK.
  • Added a comparison that changes thinking mode, temperature, and token budget one at a time in Table 31 and Appendix Section N.4, as suggested by Reviewers xRpV and LSpz.
  • Added a held-out budget sweep that reports the number of complete pairs and the error over all pairs in Table 32 and Appendix Section N.5, as suggested by Reviewer xRpV.
  • Added the schedule GAR of the model and of the exact filter under the original denominator and a denominator matched in innovation energy in Table 33 and Appendix Section N.6, as suggested by Reviewer xRpV.

Other Changes & Modifications

  • We have stated the new insight of the fixed-history audit in one sentence in Section 1, in response to Reviewer geJK.
  • We have summarized the additional models and the nonlinear task in Section 3 and the reasoning results in Section 5, in response to Reviewers xRpV, LSpz, and geJK.
  • We have replaced the development-set budget sweep in Section 5 and Appendix Section B.5 with the held-out sweep, in response to Reviewer xRpV.
  • We have clarified in Section 5 what the steering and supplied-state results establish, in response to Reviewer xRpV.
  • We have linked Appendix Sections B.3 and D.3 to Table 31 and Table 33, and added the proceedings DOI to the Liu et al. reference, in response to Reviewer xRpV.
AuthorsOpenReview ↗

Summary of the Rebuttal

Dear Reviewers and Area Chair,

Thank you for the time and effort you have devoted to evaluating our submission. We are encouraged by your comments on the strengths of our work:

  • A clear question with a well-controlled diagnostic. The fixed-history audit perturbs only the current reading and directly measures the weight that a model gives to it. (geJK, LSpz, xRpV)
  • An important distinction between state tracking and reliability-aware updating. A model can track the state accurately while the weight it gives to a new reading barely changes with the reliability of that reading. (LSpz, geJK)
  • A diagnosis connected to concrete improvements. On held-out trajectories, enabling reasoning reduces the weight error of Qwen3-8B, the mean absolute difference between its weight and the optimal weight, from 0.301 to 0.019. (geJK, xRpV)

Building on this positive feedback, we also thank the reviewers for their constructive suggestions, which we have addressed as follows:

Reviewer(s)Main concernResponse and supporting evidence
xRpV, LSpz, geJKGenerality across model families and larger modelsExtended the audit to eight models from four families, from 1.7B to 32B. As the noise increases, each adjusts the weight by less than 8% of the optimal amount (Table 28).
xRpV, LSpzGeneralization to nonlinear dynamicsAdded a nonlinear task. The optimal weight decreases from 0.515 to 0.072, whereas the weight of Qwen3-8B remains between 0.82 and 0.84 (Table 29).
xRpV, LSpz, geJKReasoning recovery beyond one modelWith reasoning enabled, the weight error of Qwen3-32B decreases from 0.179 to 0.017 (Table 30).
xRpV, LSpzSeparating thinking mode, temperature, and token budgetThe weight error decreases to about 0.02 only with thinking and a sufficient token budget. Temperature has little effect (Table 31).
xRpVBudget sweep that reports all pairsWith failed pairs counted, all 432 pairs complete at 16,384 tokens or more, with a weight error of 0.011 to 0.019 (Table 32).
xRpVSchedule GAR under a matched denominatorThe GAR of the model changes by at most 0.004, and no conclusion changes (Table 33).
geJK, xRpVNew insight relative to prior studiesStated in Section 1: accurate tracking does not imply that the weight adapts to reliability.

We are encouraged by the decision of Reviewer LSpz to raise the rating from 2 to 3, and by the reply of Reviewer geJK that "the expanded evaluation across models and reasoning settings addresses my main concerns". Reviewer xRpV has not yet responded. The requests of Reviewer xRpV in the first four rows were also raised by Reviewer LSpz, and every weakness and question has a point-by-point response in the corresponding thread.

Finally, the additional evidence shows that the finding holds across model families, model sizes, and a nonlinear task. Thank you again for your consideration and for guiding the review process.

Regards from the authors.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.