Organizations: The University of Tokyo · RWTH Aachen University · Harbin Institute of Technology, Shenzhen · The Hong Kong University of Science and Technology
Memory summaries, prompts and steering vectors set a language model's beliefs and are judged by the prediction they produce. A prediction does not say how firmly it is held. We separate a belief's content, what it predicts now, from its weight, the number of new observations it is worth: for an ideal observer, the amount of evidence behind it. In Bernoulli inference we set beliefs in nine open-weight models of 8 to 397B parameters and follow each belief with the same new observations. We argue that these models hold an answer without its evidence and recover its weight from the text they re-read. Weight therefore follows the form of the evidence and the question asked more than its amount: matched on their current estimate, a record of observations and a summary of its counts usually move apart under the same new observations, in either direction. Scale does not close the gap. Across the nine models 64 observations are worth 20 to 53 as a record and 9 to 151 as a summary, no closer to 64 at 397B than at 8B. Where no text stands behind a belief there is little to re-read: a steering vector can install a record's answer, and sixteen observations undo it where a neutral passage leaves it in place. Where reasoning first writes the counts into the context the weight returns, and the same 64 observations are worth 58 to 67 in either form. A matched answer is not a matched belief, and methods that set beliefs should be judged by the futures they produce.
Figures & tables
Figure 1: A schematic of the protocol: a belief has a content that current readouts show and a weight that only an update reveals. I. Six routes aim at the same prediction, 0.55 for heads. II. The prepared state is read as a continuation and as an estimate, and the route decides which readout shows the content. III. The same sixteen new observations follow every state, and how far each state moves gives the weight of its belief in observations (Bayes: w=n ). Values are illustrative, taken from the 8 to 12B models.
Routes
content line, proportion π∈{0.25,0.5,0.75} , 0.05 and 0.95 added for Told , Instructed
Ideal observer
Shown
The complete observations in their recorded order are: ⟨n outcomes ⟩
n , with n∈{8,16,32,64}
Summarized
A complete record contains ⟨k⟩ H outcomes and ⟨n−k⟩ T outcomes, for ⟨n⟩ observations in total.
n
Told
On independently checked previous forecasting tasks, this source's reports were judged accurate ⟨ 95 or 55 ⟩ % of the time. A source reports that the probability of H on the next trial is ⟨s⟩ .
discounted by the reliability, dropped once the evidence contradicts it (Appendix B )
Instructed
Assume that the probability of H for this coin is ⟨s⟩ . at three strengths, up to For this task, treat … as exactly ⟨s⟩ , regardless of any observations.
∞ , a constraint
Steered
a Shown state plus αv , v the mean difference between records with 75% and 25% heads, at a quarter or a half of the depth, written once at the last 4 or 16 record positions (one-shot) or kept in place at every position (persistent)
as Shown , the same content
Transplanted
residual stream at the last positions (frontier) of a Shown record of 64, copied into one of 8
as the donor, 64
Table 1: The design: six routes to a belief, two readouts with a one-token decision check, and four futures. Each row gives what the model reads (literal text on grey , generated parts in ⟨ angle brackets ⟩ , a line break) and, for a route, the weight in observations that a Bayesian observer with a uniform prior gives its content. USER and ASSISTANT mark the chat turns that define the two frames (Section 3 ). Every element is in Appendix P .
Figure 2: A statement moves the estimate and a record the continuation, and in the larger models a record moves the estimate as well. Slope of each readout on the proportion conveyed by the route (1: follows it), with 95% family-bootstrap intervals. Columns: the continuation and the estimate in the document (doc.) and assistant (asst.) frames, and the one-token decision, where a slope of 2 is an answer that flips between T and H ( Jang et al., 2026 ) . Heavy frame: the model’s sampling channel. Gemma-3-12B’s continuations at n=8 run against the record (values printed).
Figure 3: Same answer, different future. Estimate after sixteen new observations against the current estimate, under the next-toss question, for a record and a summary of 64 observations. Solid: after sixteen observations with 25% heads. Hollow: with 75% heads. The shaded gap between the two is how far a route moves. Grey: a Bayesian observer whose estimate is worth 64 observations. † : some answers after a summary are counts of the new segment, which lowers their level and leaves the gap unchanged (Appendix J , with all routes and readouts).
Figure 4: What a belief is worth in the report channel, by route, model and question. Weight w in observations under sixteen new observations (filled: expected-count question, hollow: next-toss question). Arrows run from n=8 to n=64 against the Bayesian values (dashed). Top band: not revised (the interval of S includes zero, or w exceeds 1,000). Bottom band: worth less than one observation. † : some answers are counts of the new segment (Appendix J ). Values in Appendix I , the largest models in Tables 2 and A10 .
Table 2: What 64 observations are worth in the nine models , on the same states (Appendix J ). Track: w relative to 64 on a log scale from 1/8 to 8, with its 95% interval (tick: Bayes, w=64 ).
Figure 5: The same one-shot write is undone in Llama-3.1-8B and Qwen3-8B and kept in Qwen3.8-27B. Retention R of the written content against the number of new observations, with the natural record of sixteen in grey. Hollow squares: retention after a neutral passage of the same length, for the two writes of four positions and the natural record. Llama-3.1-8B’s two written values, 1.56 and 1.66, lie in the top band. Family-bootstrap 95% intervals.
Appendix figures & tables28 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A1: Control experiments for the channel map. I. Copy control. II. Count, then estimate (with the exact-count accuracy). III. Prose-list recency. IV. Bare-stream continuation. V. Qwen3-8B in thinking mode, 1,536 against 4,096 reasoning tokens, with the non-thinking readout as baseline. VI. The five stated levels. Llama-3.1-8B, Qwen3-8B and Gemma-3-12B. In II, Gemma-3-12B’s assistant-frame point after counting at n=64 is unreadable (mass 0.04).
Figure A2: Layout dependence in the bare stream and Gemma-3-12B’s response to new evidence. I, II. Gemma-3-12B: the slope S on the new evidence under each layout (segment: a new segment; line: the extended line) in the document and assistant frames, at n=8 (hollow) and 64 (filled). III, IV. Bare 0/1 stream: S and a for the three smaller models with the new evidence on the same line (hollow) or on a new line (filled).
Figure A3: A copied frontier moves the recipient little or not at all toward the donor. Shift of the recipient toward the donor, relative to sham, for same- π and different- π donors with one and four copied positions. I. Llama-3.1-8B. II. Qwen3-8B. III. Qwen3.8-27B. The shift quoted in the text is the difference between the two donor conditions. Family-bootstrap 95% intervals.
Quarter depth
Half depth
Natural
K4
K16
K4
K16
record
Llama-3.1-8B , layers 8 and 16, document-frame continuation
Installed fraction
0.97 [0.82, 1.14]
0.92 [0.78, 1.10]
0.97 [0.80, 1.16]
1.09 [0.95, 1.29]
1
R after 8
0.24 [0.20, 0.27]
0.18 [0.14, 0.21]
0.26 [0.23, 0.30]
0.18 [0.15, 0.21]
0.55 [0.46, 0.65]
R after 16
0.07 [0.04, 0.09]
− 0.05 [ − 0.10, − 0.01]
0.12 [0.09, 0.15]
0.04 [0.01, 0.07]
0.42 [0.31, 0.53]
R after 32
0.04 [0.01, 0.07]
− 0.01 [ − 0.05, 0.03]
0.07 [0.04, 0.10]
0.06 [0.03, 0.08]
0.43 [0.32, 0.54]
Appendix
Table A1: One-shot writes against the natural record of sixteen, in the three models whose sampling channel carries the record. Installed fraction of the natural content difference, retention R after 8, 16 and 32 new observations, sensitivity S to a segment of sixteen, retention after a neutral passage in place of the segment, and the difference of the estimate between the two signs of the vector (for the natural record, between records with 75% and 25% heads). K: positions written. Family-bootstrap 95% intervals.
Figure A4: Activation writes in Llama-3.1-8B, Qwen3-8B and Qwen3.8-27B. One row per model. I. Sensitivity to a new segment of sixteen observations against the retention of the old content, for the natural record, the one-shot writes (numbers: the layer of the write) and the vector kept in place, which lies beyond the axis in Qwen3.8-27B and has no defined retention in Qwen3-8B. II. The vector kept in place at every position, as an offset on the log-odds against sham, by the number of new observations. In Qwen3-8B its negative sign fails the validity check of a write and is not drawn. Family-bootstrap 95% intervals.
m=8
m=16
m=32
Extended line
Model
n
w
R
w
R
w
R
w
R
Gemma-3-27B
8
S<0
0.84
S<0
0.12
17 [2.0, 31]
0.06
1.9 [ − 2.6, 6.8]
0.02
16
∞ ( ≥ 1227)
–
34 [27, 44]
–
13 [5.6, 20]
–
2.9 [ − 3.5, 8.5]
–
32
90 [44, 433]
8.39
49 [40, 65]
7.62
26 [19, 31]
4.20
4.6 [0.0, 9.0]
0.82
64
∞ ( ≥ 220)
2.43
114 [85, 159]
2.15
69 [58, 82]
1.64
20 [16, 25]
0.74
Qwen3.8-27B
8
1.9 [1.7, 2.2]
0.30
2.5 [2.1, 2.8]
0.21
2.8 [2.2, 3.4]
0.13
11 [10, 12]
0.49
Appendix
Table A2: A shown record in the sampling channel of Gemma-3-27B and Qwen3.8-27B. Weight w and retention R by the length n of the record and the length m of the new segment, and under the extended line of sixteen outcomes. Gemma-3-27B’s channel is close to deterministic. Family-bootstrap 95% intervals.
Figure A5: The content of every route under the same new evidence. Retention R=a/a0 of the old content after a new segment of m observations. Columns: the six models. Top: the continuation in the model’s sampling channel. Bottom: the estimate in its report channel. Dashed: the Bayesian observer for records of 8 and 64. Pink: a one-shot write at half the depth, with the natural record of the written states in grey. Top band: R above 1.1, in cells whose content is hardly read before the segment ( a0<0.35 ). Gemma-3-12B and Qwen3-32B have no sampling channel, and Gemma-3-27B’s is close to deterministic. Qwen3-8B’s sampling channel pools the two sources of a told rate. Family-bootstrap 95% intervals.
Figure A6: Weights against their Bayesian references. I. A shown record in the sampling channel against its length (new segment, m=16 ), with the Bayesian line w=n . II. The same record at n=64 against the length of the new segment, and under the extended line. I and II show Llama-3.1-8B, Qwen3-8B and Qwen3.8-27B. III. A summary in the report channel against the count it states. IV. A told rate in the report channel by source reliability. III and IV show the six models. One log scale throughout; top band: not revised. Family-bootstrap 95% intervals.
Figure A7: The sampling-channel weight of a shown record under every layout. Llama-3.1-8B, Qwen3-8B and Qwen3.8-27B. I. New segment ( m=16 ): w against the record length n , with the Bayesian line w=n . Llama-3.1-8B rises and flattens beyond n=32 , Qwen3-8B stays flat, and Qwen3.8-27B stays near a third of the record’s length. II. Extended line: markers at the top edge are not revised (primacy); black-edged markers, the strong contradiction. III. A neutral paragraph before the record: ratio of w to the undiluted record. IV. w against the segment length m at n=64 . Family-bootstrap 95% intervals.
Figure A8: The weight of every route in the estimate. Weight w of each route in the estimate of each of the six models, in the document frame and the assistant frame, after a new segment of sixteen observations. Hollow markers in the top band: not revised; crosses: moves against the evidence ( S<0 ); hollow down-triangle: content not carried ( a≈0 ); arrows: values above the axis; other hollow markers: answer mass below 0.3; grey ticks: the Bayesian w=n for summaries and records. Family-bootstrap 95% intervals.
Model
Parameters
Attention
Source
Results
Open weights, run locally: token probabilities of both readouts, activations written
Llama-3.1-8B
8B
attention-only
meta-llama/Llama-3.1-8B-Instruct
Sections 4 to 6
Qwen3-8B
8B
attention-only
Qwen/Qwen3-8B
Sections 4 to 6
Gemma-3-12B
12B
attention-only
google/gemma-3-12b-it
Sections 4 , 5
Qwen3-32B
32B
attention-only
Qwen/Qwen3-32B
Sections 4 , 5
Gemma-3-27B
27B
attention-only
google/gemma-3-27b-it
Sections 4 , 5
Appendix
Table A3: The models and how each is read. One protocol, read as deeply as each model can be accessed. For the models run locally, the repository name links to the revision that was read.
Continuation
Estimate
Decision
Model
doc.
asst.
doc.
asst.
check
Llama-3.1-8B
0.16
0.64
2.2
0.7
0.25
Qwen3-8B
0.29
0.36
8.4
5.1
0.84
Gemma-3-12B
0.59
1.15
1.1
2.7
1.46
Appendix
Table A4: Wording noise floor of the smaller models. Mean absolute difference between the wordings of the same shown record, per readout and frame: the continuation and the one-token decision check in logit units, the estimate in expected count.
Figure A9: Wording noise floor of the smaller models. Mean absolute difference between the four wordings of the same shown record, per readout, model and record length, in probability units for the continuations and the decision check and in expected count over 100 for the estimates.
Model
Route
Q1
Q2
Q3
Q4
w (Q1)
w (Q2)
Llama-3.1-8B
Told
0.99
0.99
0.99
0.97
∞ ( ≥ 13 102)
54 [40, 73]
Instructed
1.00
1.00
1.00
n.r.
∞ ( ≥ 11 987)
49 [28, 112]
Summarized, n=8
0.53
0.94
0.75
0.46
71 [35, 132]
24 [19, 29]
Summarized, n=64
0.51
0.93
0.72
0.40
52 [25, 103]
19 [14, 25]
Shown, n=8
0.12
0.42
0.62
0.20
17 [12, 23]
7.3 [5.3, 9.3]
Shown, n=64
0.02
0.43
0.15
0.06
19 [13, 27]
17 [12, 23]
Appendix
Table A5: The estimate under four queries, assistant frame. Slope of the current estimate on the content of the route (Told at the central three levels; Instructed at the central three under Q1 and at all five under Q2 to Q4, which round alike), and the weight w of the content under a new segment of sixteen observations with Q1 and with Q2, Told at the central three levels under both. Family-bootstrap 95% intervals.
Figure A10: The estimate after sixteen new observations against the current estimate, route by route. Rows: the six models. Columns: the sampling channel, the expected-count question and the next-toss question. Each point is the mean over the families of one content level, solid after sixteen observations with 25% heads and hollow after sixteen with 75% heads. Grey: a Bayesian observer whose current estimate is worth 16 or 64 observations. Dotted: no update. A route is drawn where its content reaches the readout, faded where the answer mass is below 0.3. The sampling-channel panel is empty for Gemma-3-12B and Qwen3-32B, which have no sampling channel, and faded for Gemma-3-27B, whose channel is close to deterministic.
Table A6: Equal current estimates, different futures. Pairs of states of two routes whose current estimates differ by less than 0.02: their number, the range of the matched estimates, the spread of each route (the estimate after sixteen observations with 75% heads minus that after 25%) with its difference x−y , and the difference x−y of the estimates after each segment. The larger models are shown wherever both routes can be read. Family-bootstrap 95% intervals, pairs recomputed on each resample.
Figure A11: Content and weight fall together as the record is hidden. Llama-3.1-8B. I. The weight of a record of 64 observations under sixteen new ones, against the number h of its first outcomes hidden from the positions after the record, dashed where it is proportional to the visible outcomes. II. The content slope a0 and the weight as ratios to h=0 . Family-bootstrap 95% intervals.
h
a0
S
w
w ratio
a0 ratio
0
0.70 [0.59, 0.80]
0.37
14.4 [11.3, 18.0]
1.00
1.00
16
0.67 [0.58, 0.76]
0.28
15.3 [12.2, 19.0]
1.06
0.97
32
0.44 [0.36, 0.53]
0.27
10.6 [7.3, 14.2]
0.73
0.64
48
0.24 [0.16, 0.32]
0.30
4.2 [1.1, 7.2]
0.29
0.34
Appendix
Table A7: Hiding the beginning of the record lowers content and weight together. Llama-3.1-8B’s document-frame continuation after a record of 64 observations whose first h outcomes are hidden from the positions after the record. a0 : slope of the current continuation on the proportion of the record. S , w : slope on the new evidence and weight under sixteen new observations. Ratios to h=0 from paired resamples. Family-bootstrap 95% intervals.
Prediction
Criterion
Qwen3-32B
Gemma-3-27B
Qwen3.8-27B
What sets the weight met in 5 of 6 cells
P5
The question changes a told rate’s weight
wQ2<wQ1
3 77 → 5.5
3 95 → 6.0
3 11 → 4.2
P7
A record and a summary of 64 with matched estimates move apart
Q2: Δspread=0
7 − 0.004 [ − 0.046, 0.031]
3 − 0.084 [ − 0.118, − 0.050]
3 + 0.066 d [0.056, 0.076]
The form of the dependence met in 3 of the 8 cells that can be evaluated
P1
A told rate enters the estimate and stays out of the document continuation
a0est>0.8 in every readable frame; ∣a0cont∣<0.15 , 0∈CI
7 a0est 0.73 e a0cont− 0.02 [ − 0.09, 0.05]
3 a0est 0.99 e a0cont− 0.07 [ − 0.45, 0.29]
— b
P2
A record enters the continuation more than the estimate
a0cont−a0est>0 at n=64
— a
— b
3 0.41 [0.37, 0.45]
Appendix
Table A8: The predictions fixed before the larger models were run, and their outcome in each larger model. Each cell gives the verdict and the number that decides it, with its family-bootstrap 95% interval in grey below.
Figure A12: Matched estimates in the largest models. As Figure 3 , from sampled answers: the estimate after sixteen new observations against the current estimate under the next-toss question, for a record and a summary of 64 observations. Each point is the mean over the eight states of one content level.
Table A9: A record against its summary in the largest models. For 64 observations under each question: the ratio of the two weights, with its position between one eighth and eight on the track (tick: equal weights); the spread of each route, the estimate after sixteen observations with 75% heads minus that after 25%, averaged over the range of current estimates that the two routes share, with the ratio of the two spreads; and the number of pairs of states whose current estimates differ by less than 0.02, with the difference of their spreads. Ratios are summary over record, and the difference is summary minus record. Family-bootstrap 95% intervals.
Table A10: The report channel of the largest models. Weight w of each route under sixteen new observations, by question, with family-bootstrap 95% intervals. Track: w of a record or a summary relative to the Bayesian w=n , on a log scale from one eighth to eight times (tick: w=n ).
Table A11: The direct reply in two open and three proprietary models. Weight w of each route under sixteen new observations, by question. Track and symbols as in Table A10 .
Table A12: A record against its summary in the direct reply. For 64 observations under each question, with the columns of Table A9 .
Table A13: The same states answered directly and after a reasoning step. Weight w of each route under sixteen new observations, under the next-toss question. The direct answers of Qwen3.8-27B are those of its local report channel. Track and symbols as in Table A10 .
Counts
Weight written
Answer
Model
Evidence
Traces
stated
pooled
median
at n
written
correct
Tokens
Qwen3.8-27B
record, 16
723
0.96
1.00
16
0.98
1.00
0.91
887
record, 64
754
0.99
0.99
64
0.89
1.00
0.79
1 926
summary, 16
726
1.00
0.98
16
0.97
1.00
0.93
852
summary, 64
743
1.00
1.00
64
0.99
1.00
0.94
833
Qwen3-32B
record, 16
112
1.00
1.00
16
0.96
1.00
0.96
1 785
Appendix
Table A14: What a reasoning trace writes after a new segment. Shares of traces that state the counts of the old evidence and that compute from the pooled counts; the weight that the written computation gives the old evidence, as its median and as the share of traces within one observation of n ; the shares of answers that are a value the trace wrote and that equal the pooled proportion of the true counts; and the median number of completion tokens.
Qwen3.8
Qwen3
Qwen3.5
Rule
27B
32B
122B
Observations alone
0.54
0.26
0.43
Mixture by accuracy
0.21
0.10
0.47
Posterior over the source
0.18
0.46
0.06
Rate as prior observations
0.05
0.17
0.01
Rate kept
0.01
0.00
0.03
Appendix
Table A15: The rule a trace applies to a told rate after a new segment , as shares of each model’s traces. Mixture by accuracy: the stated rate and the observations mixed by the stated accuracy of the source. Posterior over the source: over a right and a wrong source.
Figure A13: Three reasoning traces of Qwen3-32B , verbatim, with omissions marked ⟨ … ⟩ . Left: the trace pools the counts of a record of 16 observations and of a new segment. Middle: the trace miscounts a record of 64 observations with 27 heads and answers from the count it wrote, where the true count gives 42. Right: the trace handles a rate of 0.25 told by a source that is accurate 95% of the time.
Large language models (LLMs) are typically deployed in multi-turn conversations, where each turn provides new evidence that should reduce epistemic uncertainty about their environment. Acting rationally then requires inferring the unobserved quantities that govern it and updating beliefs about them as evidence accumulates. Yet most evaluations only score the model's final-turn answer in a single-turn format, leaving this process unexamined. We ask how closely LLMs' belief updates match those of a rational Bayesian reasoner in multi-turn settings, and introduce BayesBench, a suite of simulation environments that probe this across three progressively complex tasks: (i) Bayesian estimation, where the model infers an unknown parameter from sequential evidence; (ii) Bayesian prediction, where the model turns inferred beliefs about a latent variable into outcome forecasts; and (iii) latent-framed Bayesian prediction, where observations are filtered through a user-persona framing, requiring joint inference over the latent state and the persona. Across seven LLMs (3B--70B), scaling improves latent inference and evidence accumulation, with updates occasionally matching the Bayesian posterior. However, these gains do not reliably carry over to downstream prediction, exposing a gap between inferring latent structure and using it to rationally update beliefs about the target outcome.
Ankur Samanta, Akshayaa Magesh, Tal Lancewicki +7
Meta AI · Columbia University · Work done at Meta +2
Modern AI systems are being deployed in complex domains such as medicine, science, and law, where it is important that they not only produce correct answers, but also represent and update uncertain beliefs about the world as new evidence arrives. We introduce the novel technique of studying LLMs as information processing rules and utilize the information processing gap to study the internal (in)consistencies of how LLMs update their probabilistic beliefs from evidence. Our extensive experiments evaluate multiple approaches in which LLMs can incorporate evidence into their beliefs. Some of these approaches produce (nearly) Bayesian updates; others seem to use a learned heuristic. Surprisingly, the non-Bayesian heuristic updates often outperform exact Bayesian computation in terms of downstream task performance -- indicating the LLMs' probabilistic models of the world are misspecified. Lastly, we show how our measure can provide diagnostics to identify issues with LLM-powered inferential systems.
Chacha Chen, Matthew Jörke, Adam Goliński +4
1Apple · 2Stanford University · †Work done while an intern at Apple. +1
Long-horizon interactions require language models to manage accumulating information: when to update their state, when to preserve their state, and what to ignore. We study this challenge as Contextual Belief Management (CBM): maintaining a predicted belief state aligned with formal evidence while isolating task-irrelevant noise. To make CBM measurable, we introduce BeliefTrack, a closed-world benchmark spanning Rule Discovery and Circuit Diagnosis, where a finite belief space and symbolic verifiers enable exact turn-level evaluation. BeliefTrack diagnoses three failures: Failed Stay, Failed Update, and Failed Isolation. Across multiple LLMs, vanilla models exhibit severe CBM failures, while explicit belief-tracking prompts provide limited gains. In contrast, reinforcement learning with belief-state rewards reduces failure rates by 70.9% on average. Further probing reveals latent belief-state dynamics behind these failures, and representation-level steering reduces failure rates by 46.1% across two tasks (Code is available at https://github.com/zjunlp/CBM).